• Suppr超能文献
  • 文献检索
  • 文档翻译
  • 深度研究
  • 学术资讯
  • Suppr Zotero 插件Zotero 插件
定价套餐&价格
应用&插件
Suppr Zotero 插件Zotero 插件浏览器插件Mac 客户端Win 客户端微信小程序
定价
会员套餐积分包API 积分包
服务
文献检索文档翻译深度研究API 文档MCP 服务
关于我们
关于 Suppr公司介绍联系我们用户协议隐私条款
关注我们

Suppr 超能文献

核心技术专利：CN118964589B侵权必究
粤ICP备2023148730 号-1Suppr @ 2026
  1. 首页
  3. 分享广场
  5. 基于机器学习的肺癌患者免疫治疗后肺间质病预测研究进展

基于机器学习的肺癌患者免疫治疗后肺间质病预测研究进展

深度研究郑建光发表于 2025年10月23日 09:553阅读
发起深度研究
发起深度研究

1. 背景与研究意义

1.1 肺癌免疫治疗与肺间质病的临床现状

肺癌是全球范围内发病率和死亡率最高的恶性肿瘤之一。近年来，以程序性死亡受体-1（PD-1）及其配体（PD-L1）抑制剂为代表的免疫检查点抑制剂（ICIs）的出现，彻底改变了晚期非小细胞肺癌（NSCLC）的治疗格局，显著延长了患者的生存期 12。对于大多数晚期NSCLC患者，PD-1或PD-L1抑制剂已成为一线治疗方案 1。然而，ICIs在激活抗肿瘤免疫反应的同时，也可能导致一系列免疫相关不良事件（irAEs），这些事件可累及全身多个器官系统 23。

在所有irAEs中，免疫相关性肺炎（CIP），也被称为ICI相关间质性肺病（ILD），是非小细胞肺癌患者中发生率较高且可能致命的并发症之一 23。尽管CIP的发生率相对较低（全级别0-10%，3-4级为0-3%），但其严重程度不容忽视，甚至可能导致死亡 45。据报道，接受抗PD-1/PD-L1单抗治疗的患者中，肺炎的发生率约为5%，而联合抗细胞毒性T淋巴细胞相关蛋白4（CTLA-4）单抗的治疗方案则将该风险提高至10% 4。对于NSCLC患者，CIP的发生率和中位发病时间因治疗方案和患者特征而异 3。

CIP的临床表现多样，可与原发性肺癌的呼吸道症状和影像学表现高度重合，这给其早期检测、诊断和管理带来了巨大挑战 23。放射学上，CIP可表现为多种模式，包括机化性肺炎（OP）、非特异性间质性肺炎（NSIP）、超敏性肺炎、急性间质性肺炎/急性呼吸窘迫综合征（ARDS）及细支气管炎等 56。值得注意的是，预先存在的肺部疾病（如间质性肺病、肺气肿）以及既往胸部放疗均是CIP发生的高风险因素 789。有效的治疗通常包括停用ICIs和使用糖皮质激素，但部分患者可能发展为类固醇难治性CIP，需要额外的免疫抑制剂治疗，预后较差 710。

由于CIP的潜在严重性及其对后续免疫治疗的负面影响（可能需要停药或调整治疗方案），早期准确预测高危患者至关重要 23。这将有助于临床医生在免疫治疗开始前进行风险分层，并制定个性化的监测和干预策略，从而优化治疗决策，降低CIP的发生率和严重程度，最终改善患者的临床结局。

1.2 机器学习在肺间质病预测中的价值

传统的肺间质病预测方法主要依赖于临床经验、生物标志物和基于统计学的风险评分模型。这些方法通常基于有限的临床指标，如患者年龄、吸烟史、合并症以及部分实验室检查结果。然而，这些传统方法存在显著局限性：首先，它们往往难以有效整合多源异构数据，例如复杂的影像学特征、基因组学数据和治疗方案细节；其次，肺间质病的发病机制复杂，涉及多种非线性相互作用，而传统模型在捕捉这些复杂关系方面表现不佳；最后，由于缺乏对个体差异的精细刻画，传统方法的预测精度和个体化风险评估能力有限，难以满足临床上对高风险患者进行早期、精准识别的需求。

近年来，随着人工智能技术，特别是机器学习（Machine Learning, ML）的飞速发展，其在医疗健康领域的应用展现出巨大潜力 11。机器学习模型能够处理海量且多维度的数据，从复杂的临床、影像和组学数据中自动学习潜在模式和非线性关系，从而在疾病预测和诊断方面实现更高的准确性 12。在肺间质病预测中，机器学习的优势体现在以下几个方面：

  1. 多源数据整合能力：机器学习能够有效地融合来自不同模态的数据，如患者人口统计学信息、临床生化指标、详细的治疗方案、医学影像（CT、PET-CT）特征以及组学数据（基因组、转录组）等 1314。这种多模态数据整合能力克服了传统方法对单一或少量数据类型依赖的局限性，为更全面的风险评估提供了基础。
  2. 非线性关系挖掘：机器学习算法在处理具有复杂非线性模式的数据方面表现出色 1516。肺间质病的发生并非由单一因素线性决定，而是多种因素相互作用的结果。机器学习能够捕捉这些微妙且复杂的相互作用，揭示传统统计方法难以发现的潜在预测因子。例如，随机森林（Random Forest, RF）和支持向量机（Support Vector Machine, SVM）等算法已被证明在处理高维非线性数据时具有优越性能 17。
  3. 预测精度提升：多项研究表明，机器学习模型在预测各种疾病（包括心血管事件、败血症相关谵妄等）方面的性能优于传统统计模型 181920。对于肺间质病，结合影像组学特征和临床数据的机器学习模型展现出显著的预测能力，一些研究报告其曲线下面积（AUC）可达到0.922，远超仅使用临床信息或放射组学特征的模型 13。这种高预测精度有助于医生更早地识别高风险患者，从而提前采取预防措施或调整治疗方案，以改善患者预后。
  4. 风险分层与决策支持：通过机器学习模型对患者进行精确的风险分层，临床医生可以根据患者个体特征制定更个性化的免疫治疗策略，例如调整免疫检查点抑制剂的剂量、联合用药方案，或加强对高风险患者的监测频次。此外，可解释性机器学习（Interpretable ML）技术如SHapley Additive exPlanations (SHAP)值的应用，能够帮助临床医生理解模型预测的依据，增强模型的可信度和临床接受度 2122。

本综述的核心目标在于系统性地回顾和分析当前基于机器学习技术预测肺癌患者免疫治疗后肺间质病的研究进展。我们将重点探讨不同数据模态（如影像、临床、组学数据）的整合策略、主流机器学习模型的构建与优化方法、模型的性能评估标准，以及其在临床应用中的潜在价值。同时，本综述也将深入剖析当前研究面临的挑战，并展望未来的研究方向，以期为肺癌免疫治疗相关肺间质病的早期预测和管理提供新的思路和科学依据。

2. 预测模型的数据来源与特征提取

2.1 多模态数据类型与采集

在基于机器学习预测肺癌患者免疫治疗后肺间质病的研究中，数据的全面性和多样性是构建高效预测模型的基石。为了捕捉肺间质病发生发展的复杂生物学过程，研究人员通常会整合多种模态的数据，包括影像数据、临床数据、组学数据以及功能数据。每种数据类型都承载着独特的生物学信息，它们相互补充，共同构成了预测模型的输入特征。

  1. 影像数据：

    • 治疗前/后CT影像：高分辨率计算机断层扫描（HRCT）是评估肺部病变，特别是间质性肺病的金标准。治疗前的CT影像可以提供肺部基线状况，包括是否存在既往间质性病变或纤维化，这些是免疫治疗后肺毒性的重要风险因素 232425。治疗中或治疗后的CT影像则可用于监测肺部变化，并提取动态特征。通过对CT影像进行分析，可以量化肺实质的密度、纹理和结构特征，例如磨玻璃影、网格影、蜂窝肺等，这些都与间质性肺病的病理改变密切相关。
    • 影像数据在预测中的生物学关联在于，它们能够直观地反映肺部的解剖学和病理生理学改变。通过影像组学（Radiomics）技术或深度学习方法，可以从这些图像中提取出人眼难以察觉的定量特征，如肿瘤或肺实质的纹理、形状、强度等，这些特征可能与免疫治疗引起的肺部炎症和纤维化过程密切相关 13。例如，肺实质的特定影像特征可能预示着免疫细胞浸润或炎症反应的强度。

  2. 临床数据：

    • 患者基本信息：包括年龄、性别、吸烟史、合并症（如慢性阻塞性肺疾病COPD、自身免疫性疾病）、既往肺部疾病史等。这些因素是肺间质病发生的已知风险因子。例如，COPD患者被发现有更高的肺间质病风险 26。
    • 免疫治疗方案：详细的治疗方案信息至关重要，包括免疫检查点抑制剂的类型（PD-1、PD-L1、CTLA-4抑制剂及其联合用药）、剂量、疗程、是否联合放疗或化疗等。不同免疫治疗方案的肺间质病发生风险存在差异。例如，有研究表明抗PD-L1单抗相比抗PD-1或抗CTLA-4单抗，可能带来更高的肺部不良事件风险 27。此外，免疫治疗与放疗的联合应用被证实会显著增加肺损伤的风险 1327。
    • 实验室指标：血液常规、肝肾功能、炎性标志物（如C反应蛋白、白细胞介素-6）、甲状腺功能指标等。甲状腺功能异常，例如甲状腺素水平，已被临床证据支持与间质性肺病相关 27。
    • 临床数据提供了患者的个体化背景信息和治疗相关因素，有助于识别易感人群和评估治疗方案对肺间质病风险的影响。

  3. 组学数据：

    • 基因组数据：通过基因组测序，可以识别与肺间质病易感性相关的遗传变异，例如免疫相关通路基因的单核苷酸多态性（SNPs）或基因表达改变。这些数据可以揭示个体对免疫治疗产生肺部毒性的遗传 predisposition。
    • 转录组数据：通过RNA测序等技术获取治疗前或治疗过程中肺组织或外周血细胞的基因表达谱。特定基因的表达水平或基因通路的变化可能作为免疫治疗相关肺间质病早期预警的生物标志物。例如，免疫相关通路中的基因表达变化可能反映了免疫应答的强度或失调，从而导致肺部炎症和损伤。
    • 组学数据提供了分子层面的信息，有助于深入理解肺间质病的生物学机制，并发现潜在的分子标志物，从而实现更精准的风险预测。

  4. 功能数据：

    • 肺功能指标：包括用力肺活量（FVC）、第一秒用力呼气容积（FEV1）及其比值、弥散功能（DLCO）等。这些指标反映了肺部的通气和换气功能，治疗前基线肺功能下降或治疗过程中肺功能进行性恶化可能提示肺间质病的发生或进展。
    • FDG-PET代谢特征：氟代脱氧葡萄糖正电子发射断层扫描（FDG-PET）可以量化肿瘤或正常组织的葡萄糖代谢活性，通常用标准化摄取值（SUV）表示。研究发现，肿瘤代谢反应（如治疗后SUVmean的显著下降）与较高的肺炎风险相关，这可能与患者的免疫原性差异有关 26。高代谢活性可能反映局部炎症或免疫激活状态。
    • 功能数据从生理学角度补充了对肺部健康的评估，尤其是在监测治疗效果和识别早期功能损伤方面具有独特优势。

综合运用这些多模态数据，机器学习模型能够建立一个更为全面和精确的预测框架，从而提高对肺癌患者免疫治疗后肺间质病的早期识别能力。

2.2 特征提取技术与筛选

在多模态数据采集的基础上，特征提取与筛选是机器学习预测模型构建的关键环节，旨在从原始数据中提炼出对预测目标（肺间质病）具有高判别力的信息，同时降低数据维度，优化模型性能。

  1. 影像特征提取

    • 放射组学（Radiomics）特征：放射组学通过从CT、PET等医学影像中提取高通量定量特征来描述肿瘤及其微环境的表型。这些特征可以分为几大类：
      • 形态学特征（Shape features）：描述区域（如肺实质、病灶）的大小和形状，例如体积、表面积、球形度、紧凑度等。
      • 一阶统计特征（First-order statistics）：描述像素或体素强度分布的统计学特征，如均值、中位数、方差、偏度、峰度等，反映图像强度的整体分布。
      • 纹理特征（Texture features）：通过分析像素或体素的空间关系来量化图像的异质性，例如灰度共生矩阵（GLCM）、灰度游程长度矩阵（GLRLM）、灰度区域大小矩阵（GLSZM）、邻域灰度差分矩阵（NGTDM）、灰度依赖矩阵（GLDM）等。这些特征可以捕捉肺实质的微观结构变化，如磨玻璃影、网格影的精细纹理，这些变化常与肺间质病的病理改变相关。例如，在预测放疗后肺炎的研究中，放射组学特征被证明具有预测价值 1328。
      • 小波变换特征（Wavelet features）：通过小波变换将图像分解成不同频率分量，从而提取多尺度纹理信息，可以更全面地捕捉图像的局部和全局特征。
        放射组学特征的提取通常依赖于特定的软件工具，如3D Slicer、Pyradiomics等，并需要对感兴趣区域（ROI）进行精确勾画，例如肺实质、肿瘤体积（GTV）、计划靶区（PTV）等 1328。
    • 深度学习特征（Deep Learning-based Radiomic, DLR）：随着深度学习技术的发展，尤其是卷积神经网络（CNN）在图像识别领域的成功，利用预训练的深度神经网络（如ResNet、VGG、Inception等）作为特征提取器已成为一种趋势。这些网络在大量自然图像数据集上学习了丰富的图像特征表示，其内部的卷积层和全局平均池化层能够自动捕捉图像的深层、抽象特征，无需人工设计。例如，有研究利用预训练的ResNet34模型从预放疗CT图像中提取深度图像特征，并结合放射组学和临床数据来预测肺间质病，取得了优异的性能 13。ResNet模型因其能够有效解决深度网络训练中的梯度消失问题，被广泛应用于医学影像特征提取 132829。这种方法提取的特征往往比传统放射组学特征具有更高的抽象度和判别力，能够更好地捕捉复杂的影像模式。

  2. 临床/剂量学特征提取

    • 临床特征：直接从患者病历中获取，包括人口统计学信息（年龄、性别）、吸烟史（包年数）、合并症、既往治疗史等。
    • 剂量学特征（Dosiomics）：主要针对接受放疗的患者，从放疗计划中提取与肺部受照剂量相关的指标。
      • 平均肺剂量（Mean Lung Dose, MLD）：全肺或非肿瘤肺组织接受的平均放射剂量，是预测放疗性肺炎的经典指标。
      • 肺体积受照剂量（Vx）：指肺部接收到高于x Gy剂量的体积百分比。常见的有V5、V10、V20、V30，分别代表肺体积接收到5 Gy、10 Gy、20 Gy、30 Gy或更高剂量的百分比。V20被认为是预测放疗性肺炎最重要的剂量学参数之一 30。
      • 剂量-体积直方图（DVH）特征：除了Vx和MLD，还可以从DVH曲线中提取更多精细的剂量学特征，如剂量均匀性指数、适形度指数等。
      • 4D-CT通气功能特征：一些研究结合4D-CT影像，评估肺部不同区域的功能性，并计算功能肺区域（functioning zones）的剂量-功能特征。例如，通过设置不同的通气阈值（如低灌注区域或高灌注区域），计算这些功能区域的平均剂量或受照体积百分比，这些特征被证明可以提高放疗性肺炎的预测能力 31。
        剂量学特征直接量化了肺组织所承受的物理损伤，与肺间质病的发生具有明确的生物学关联。

  3. 组学特征提取

    • 基因组特征：主要通过基因测序（如全基因组测序、外显子测序）识别与疾病风险相关的单核苷酸多态性（SNPs）、拷贝数变异（CNVs）或基因表达量。
    • 转录组特征：通过RNA测序或基因芯片技术，获取特定基因或基因集的表达水平。这些数据可以通过生物信息学工具进行差异表达基因分析、通路富集分析，以识别与肺间质病发生发展相关的分子标志物或通路。

  4. 特征筛选策略
    特征数量的庞大可能导致模型过拟合和计算效率下降，因此，有效的特征筛选至关重要。

    • 统计学方法：
      • t检验/卡方检验：用于评估单个特征与目标变量之间的统计学显著性，筛选出具有显著差异的特征。
      • 相关性分析：计算特征之间的相关系数，去除高度冗余的特征，以减少多重共线性。
    • 基于模型的选择方法：
      • LASSO（Least Absolute Shrinkage and Selection Operator）回归：LASSO是一种L1正则化方法，在模型训练过程中可以同时进行特征选择和正则化。它通过惩罚系数的绝对值和，使得不重要的特征的系数趋向于零，从而实现特征选择 313233。LASSO在医学领域，尤其是在处理高维数据时，因其出色的特征选择能力而被广泛应用。
      • 随机森林（Random Forest, RF）重要性排序：随机森林算法在构建决策树集成时，可以评估每个特征对模型预测能力的贡献度，从而给出特征的重要性得分 30。研究人员可以根据这些得分选择最重要的特征子集。
      • 递归特征消除（Recursive Feature Elimination, RFE）：这是一种贪婪的优化算法，它通过递归地训练模型并去除最不重要的特征来选择特征子集。
    • 降维方法：
      • 主成分分析（Principal Component Analysis, PCA）：将原始高维特征空间映射到较低维度的空间，生成新的主成分特征，这些主成分是原始特征的线性组合，能够保留大部分数据方差。
        有效的特征筛选不仅可以提高模型的泛化能力，避免过拟合，还能增强模型的可解释性，帮助临床医生理解哪些因素对肺间质病风险预测最为关键。例如，在预测放疗性肺炎的研究中，结合临床、剂量学和放射组学特征，通过LASSO等方法筛选出关键特征，能够显著提升模型的预测性能 33。

3. 机器学习模型的构建与性能评估

在肺癌患者免疫治疗后肺间质病（CIP）的预测研究中，机器学习模型的选择和构建是决定预测性能的关键环节。研究者们探索了多种算法和模型架构，以期从多模态数据中捕捉复杂的模式，实现精准预测。这些模型大致可分为传统机器学习算法和深度学习模型两大类，它们在处理不同类型数据和提取特征方面各有侧重。

3.1 主流算法与模型架构

  1. 传统机器学习算法
    传统机器学习算法通常需要预先进行特征工程，即人工或半人工地提取数据中的有效特征，然后再输入模型进行学习和分类。在肺间质病预测中，常见的传统算法包括：

    • 随机森林（Random Forest, RF）：随机森林是一种集成学习方法，通过构建多棵决策树并汇总它们的预测结果来提高模型的准确性和稳定性。RF具有处理高维数据、非线性关系和较少过拟合的优点，并且能够评估特征的重要性，这对于理解哪些临床或影像特征对CIP预测贡献最大非常有价值。随机森林是机器学习研究中常用的方法之一 34。
    • 支持向量机（Support Vector Machine, SVM）：SVM是一种经典的二分类算法，通过在高维空间中找到一个最优超平面，将不同类别的样本最大化地分开。支持向量机是一种常用的机器学习技术 34，有研究利用SVM结合放射组学特征预测肺癌预后，取得了较好的效果 35。
    • 梯度提升决策树（Gradient Boosting Decision Tree, GBDT）/XGBoost：这些算法也是集成学习的代表，通过迭代地训练弱学习器（通常是决策树）并优化损失函数来构建强预测模型。XGBoost是一种机器学习方法 34。
    • 逻辑回归（Logistic Regression）：作为一种广义线性模型，逻辑回归虽然相对简单，但其结果易于解释，常作为基线模型或与其他复杂模型结合使用，尤其适用于临床指标的二分类预测。

  2. 深度学习模型
    深度学习，特别是卷积神经网络（Convolutional Neural Network, CNN），在处理图像、序列等复杂非结构化数据方面展现出强大能力，能够自动从原始数据中学习和提取多层次、抽象的特征，无需繁琐的人工特征工程。

    • 卷积神经网络（CNN）：CNN在医学影像分析领域具有显著优势，特别适用于CT影像特征的提取和分类 36。其核心在于卷积层能够通过滤波器自动学习图像的局部特征（如边缘、纹理），池化层则用于降采样和特征聚合，从而捕捉图像的空间层级信息。
      • 2D CNN：适用于单张CT切片的分析，能够有效识别肺部病变的纹理、密度等特征。例如，在肺结节的良恶性分类中，CNN能够从CT图像中提取深层特征，达到较高的诊断准确率 3738。
      • 3D CNN：对于三维医学影像数据（如胸部CT扫描），3D CNN能够更好地捕捉相邻切片之间的空间信息，从而更全面地理解肺部结构和病灶特征。例如，3D U-Net等深度学习模型已被成功应用于肺部病变的自动分割，这为后续的特征提取和分析奠定了基础 3940。在CIP预测中，3D CNN可以直接从原始CT图像中学习与肺间质病相关的深层影像模式。
    • 残差网络（ResNet）：ResNet通过引入残差连接（skip connections），有效解决了深度神经网络训练过程中梯度消失和模型退化的问题，使得网络可以构建得非常深，从而学习到更复杂的特征表示。ResNet系列模型（如ResNet-50）在医学图像分类、分割和特征提取任务中表现出色，被广泛用于从CT图像中提取深度影像特征，以预测疾病进展和预后 414243。
    • 其他深度学习架构：
      • 循环神经网络（Recurrent Neural Network, RNN）及其变体（如LSTM）：虽然主要用于序列数据，但在处理时间序列的临床数据或动态影像数据时可能具有潜力，但目前在CIP预测中应用较少。
      • Transformer：近年来在自然语言处理和计算机视觉领域大放异彩，其自注意力机制使其能够捕捉长距离依赖关系，未来可能应用于多模态医学数据的融合和分析 3644。

模型选择与数据类型的适配性：
模型选择并非一概而论，需要根据数据类型和研究目标进行适配。

  • 对于以结构化临床数据、剂量学数据为主的预测任务，传统机器学习算法如随机森林、XGBoost通常表现良好，并且具有较好的可解释性。这些模型能够有效地处理数值型和类别型特征，并识别出不同特征对预测结果的贡献度。
  • 对于以医学影像数据（如CT）为主的预测任务，深度学习模型，特别是CNN及其变体（如ResNet），展现出压倒性优势。CNN能够自动从像素层面学习具有判别力的特征，避免了人工设计放射组学特征的复杂性和潜在的主观性。卷积神经网络（CNN）在图像分析中被广泛使用，能够有效识别和提取图像特征。
  • 当面对多模态数据时，可以将不同模态的特征（例如，CNN从CT中提取的深度特征、放射组学特征、临床指标、组学数据）进行融合，然后输入到传统机器学习模型或设计特定的多模态深度学习架构进行预测。多模态融合策略将在下一节详细讨论。

选择合适的模型架构需要综合考虑数据量、数据类型、计算资源、模型的复杂度和可解释性要求。例如，对于数据量较小的情况，传统机器学习算法或迁移学习（利用预训练深度模型进行特征提取）可能更为合适。随着数据量的增加和计算能力的提升，端到端的深度学习模型将展现出更强的潜力。

3.2 多模态融合策略与优化

在肺间质病的预测中，单一模态数据往往无法全面捕捉疾病发生的复杂机制。因此，将不同来源的数据（如影像、临床、组学数据）进行有效融合，构建多模态预测模型，是提升预测性能的重要途径。多模态融合策略通常分为早期融合（Early Fusion）和晚期融合（Late Fusion）两种主要方式 45。

  1. 早期融合（Early Fusion / Feature-level Fusion）
    早期融合是指在模型输入阶段，将所有模态的原始特征或提取出的低级特征直接拼接成一个单一的、高维的特征向量，然后输入到一个机器学习模型中进行训练。这种方法假设不同模态的数据特征在输入层就相互关联和补充。

    • 优点：
      • 能够捕捉到不同模态特征之间的潜在相互作用，因为所有特征都在同一个特征空间中进行学习。
      • 模型结构相对简单，仅需训练一个模型。
    • 缺点：
      • 对特征对齐和标准化要求较高，不同模态的数据量纲、噪声水平可能差异很大。
      • 高维特征向量可能导致“维度灾难”，增加模型训练难度和过拟合风险，尤其是在数据量有限的情况下。
    • 应用案例：在预测肺栓塞（PE）的研究中，有方法通过将高维图像特征和电子病历（EMR）特征进行选择性融合后，输入到机器学习模型中，实现了0.947的AUROC 45。类似地，在肺间质病预测中，可以将从CT影像中提取的放射组学特征、深度学习特征，与临床指标（如年龄、吸烟史、既往病史等）、实验室数据等直接合并，形成一个综合特征向量，再由如随机森林、SVM或神经网络进行分类预测。

  2. 晚期融合（Late Fusion / Decision-level Fusion）
    晚期融合是指对每个模态的数据分别训练独立的预测模型，然后将各个模型的预测结果（如概率、类别标签）在决策层进行组合。组合方法可以是简单的投票、加权平均、堆叠（Stacking）或其他集成学习策略。

    • 优点：
      • 允许针对不同模态数据的特性选择最合适的模型架构和训练策略。
      • 当某一模态数据缺失时，其他模态的模型仍可独立运行。
      • 降低了单一高维特征向量带来的维度灾难风险。
    • 缺点：
      • 无法直接捕捉到模态内部特征之间的细粒度交互。
      • 需要训练多个模型，计算成本可能较高。
    • 应用案例：在帕金森病识别的研究中，一种自适应集成堆叠（AE_Stacking）模型通过协同整合基于MRI和遗传数据训练的多个基础分类器的预测结果，实现了95.36%的平衡准确率和0.974的AUC，显著优于特征级融合和单模态模型 46。对于肺间质病，可以分别训练一个基于CT影像的深度学习模型、一个基于临床指标的传统机器学习模型，然后通过加权投票或另一个元学习器（meta-learner）来整合它们的预测结果。

  3. 混合融合（Hybrid Fusion）
    混合融合结合了早期融合和晚期融合的特点，例如，可以在特征提取阶段对不同模态数据进行初步融合，或者在决策层融合的基础上，再引入一些跨模态的特征交互层。深度学习模型在设计时可以天然地实现混合融合，例如，通过在网络的不同层引入不同模态的特征，并利用注意力机制等技术实现模态间的交互。

模型优化技术

为了提升模型的泛化能力和预测精度，多种优化技术被广泛应用：

  • 交叉验证（Cross-validation）：特别是五折交叉验证（5-fold cross-validation）或十折交叉验证，是评估模型性能和选择超参数的常用方法。它将数据集划分为k个子集，轮流用k-1个子集训练模型，用剩下的1个子集进行测试，重复k次，最终将k次结果的平均值作为模型的性能评估。这有助于减少模型对特定训练集过拟合的风险，提供更稳健的性能评估。
  • 数据增强（Data Augmentation）：尤其在医学影像领域，由于高质量标注数据获取困难且样本量有限，数据增强成为提高模型泛化能力的关键技术 4748。通过对现有数据进行各种变换，如旋转、平移、缩放、翻转、裁剪、添加噪声、弹性形变、颜色抖动等，可以生成新的训练样本，从而扩大训练数据集的规模和多样性，减少过拟合 495051。例如，在CT图像中，可以通过随机旋转、亮度调整等方式生成新的肺部图像，使模型在更广泛的图像变异下仍能保持鲁棒性。有研究通过数据增强技术成功解决了机器学习模型随时间推移的“模型漂移”问题，并通过增加新数据训练了新的随机森林模型，显著提高了召回率 52。
  • 超参数调优（Hyperparameter Tuning）：通过网格搜索（Grid Search）、随机搜索（Random Search）或贝叶斯优化等方法，寻找模型的最优超参数组合，以最大化模型在验证集上的性能。
  • 正则化（Regularization）：如L1、L2正则化、Dropout等，通过对模型复杂度施加惩罚，防止模型过拟合训练数据。

多模态融合的性能提升案例

多模态融合模型在肺间质病预测中展现出显著的优势。研究表明，结合了影像（如放射组学特征、深度学习特征）、临床和组学数据的多模态模型，其预测性能通常优于仅依赖单一模态数据的模型。例如：

  • 有研究发现，在预测肺部疾病时，结合CT影像和临床电子病历数据的多模态融合模型能够显著提高诊断准确性 45。
  • 在其他癌症预测任务中，整合多种模态（如超声图像、临床记录）的深度学习网络在预测淋巴结转移方面，其F1分数可达0.888，AUC值达到0.973，显著优于三个单模态深度学习网络 53。这表明多模态信息能够提供更全面的视角，从而做出更精准的预测。
  • 对于肺间质病的预测，一些研究报告单模态模型（如仅使用放射组学或仅使用临床数据）的AUC可能在0.7-0.8之间。例如，一项预测脑动静脉畸形放疗结果的研究中，最佳预测模型的平均AUC为0.71 54；另一项预测退行性颈椎脊髓病术后结果的研究中，最佳预测模型的平均AUC为0.70 55。然而，当将影像特征与临床数据进行融合后，模型的性能可以得到显著提升，例如在预测透明细胞肾细胞癌PBRM1突变状态的研究中，随机森林算法使用CT纹理分析特征的AUC值达到了0.987 56。这种性能提升验证了多模态融合策略在捕获疾病复杂性方面的有效性，它允许模型整合不同层次的信息，从而更准确地识别高风险患者。

3.3 预测性能的评估指标

为了客观、全面地衡量机器学习模型在预测肺癌患者免疫治疗后肺间质病方面的效能，需要采用一系列标准化的评估指标。这些指标不仅能反映模型的准确性，还能体现其在临床应用中的实用价值。

  1. 受试者工作特征曲线下面积（Area Under the Receiver Operating Characteristic Curve, AUC）：
    AUC是评估二分类模型预测性能最常用的指标之一，它综合反映了模型在所有可能分类阈值下的真阳性率（敏感性）和假阳性率（1-特异性）之间的权衡关系。AUC值越接近1，表明模型的分类性能越好；AUC值为0.5则表示模型性能与随机猜测无异。在肺间质病预测研究中，高AUC值意味着模型能够更好地将发生肺间质病的患者与未发生肺间质病的患者区分开来。
    例如，有研究 13开发的基于深度学习的多模态融合模型（结合深度图像特征、放射组学和临床数据）在预测免疫治疗合并放疗后≥2级症状性肺炎（Symptomatic Pneumonitis, SP）时，展示了0.922的AUC（95% CI: 0.902-0.945），远高于仅使用放射组学特征的模型（AUC 0.811）或仅使用临床信息模型（AUC 0.711）。这表明多模态融合能够显著提升预测性能。其他研究也报告了类似的优异表现，如某些模型对≥2级肺间质病的预测AUC范围为0.82-0.94，进一步印证了机器学习模型在这一领域的巨大潜力。

  2. 敏感性（Sensitivity / Recall）：
    敏感性是指模型正确识别出所有实际发生肺间质病患者的比例。高敏感性对于肺间质病这类潜在严重并发症的预测至关重要，因为它能确保高危患者尽可能不被漏诊，从而争取早期干预的机会。例如，一项预测髋骨骨质疏松的机器学习模型分析显示，其汇总敏感性为0.844 57，而另一项预测骨质疏松的深度学习模型的汇总敏感性为0.86 58。在肺间质病预测中，这意味着在所有实际发生肺间质病的患者中，有84.4%或86%的患者被模型正确预测出来。

  3. 特异性（Specificity）：
    特异性是指模型正确识别出所有实际未发生肺间质病患者的比例。高特异性有助于减少误诊，避免对低风险患者进行不必要的过度监测或干预，从而减轻医疗负担和患者焦虑。例如，上述骨质疏松预测研究中，机器学习模型的汇总特异性为0.781 57，深度学习模型的汇总特异性为0.89 58。这意味着在所有未发生肺间质病的患者中，有78.1%或89%的患者被模型正确预测出来。

  4. 诊断优势比（Diagnostic Odds Ratio, DOR）：
    DOR是评估诊断测试准确性的另一个综合指标，它结合了敏感性和特异性。DOR定义为真阳性率与假阴性率之比除以假阳性率与真阴性率之比，或简单地说是阳性似然比除以阴性似然比。DOR值越高，表示模型的诊断性能越好。DOR值的范围从0到无穷大，理论上，DOR为1表示模型没有鉴别能力。例如，在预测骨质疏松的研究中，深度学习模型的DOR高达49.09 58，远高于传统机器学习模型的18.91 57，显示了深度学习在诊断准确性上的优势。在COVID-19的诊断中，深度学习的DOR达到了112.5 59。

  5. 准确性（Accuracy）：
    准确性是指模型正确预测的样本数占总样本数的比例，是最直观的评估指标。然而，当正负样本数量不平衡时，准确性可能具有误导性，因此通常需要结合其他指标进行评估。例如，预测糖尿病患者截肢风险的机器学习模型，其测试数据的准确性为77.8% 60。

  6. 阳性预测值（Positive Predictive Value, PPV）和阴性预测值（Negative Predictive Value, NPV）：
    PPV是指被模型预测为阳性的患者中，实际真正发生肺间质病的比例。NPV是指被模型预测为阴性的患者中，实际真正未发生肺间质病的比例。这两个指标在临床实践中尤其重要，因为它们直接回答了“如果测试结果为阳性/阴性，那么患者患病/不患病的概率是多少”的问题，对临床决策具有指导意义。

  7. F1分数（F1-score）：
    F1分数是敏感性（召回率）和精确率（PPV）的调和平均值，特别适用于处理类别不平衡的数据集。它提供了一个平衡的度量，避免了仅关注某一类别的性能。

模型在不同风险分层中的表现差异：
研究表明，机器学习模型在预测不同级别的肺间质病（如≥2级或≥3级）时可能表现出性能差异。通常，预测更严重的并发症（例如3-4级，需要更积极干预）时，由于样本量较少，模型训练的难度可能增加，但在临床上，对这些高危事件的精准预测价值更大。通过调整分类阈值，可以在敏感性和特异性之间进行权衡，以适应不同的临床需求，例如，在高危环境中选择更高的敏感性以避免漏诊。

总而言之，多样的评估指标共同构成了对机器学习预测模型性能的全面画像。在肺癌免疫治疗后肺间质病的预测中，综合运用这些指标，并结合临床实际情况进行分析，能够更准确地评估模型的临床应用价值，并指导其进一步的优化和转化。

4. 临床应用场景与实践价值

4.1 治疗前风险分层与决策支持

在肺癌免疫治疗领域，如何精准识别免疫治疗后肺间质病（CIP）的高风险患者，并在治疗前进行有效的风险分层，是优化治疗策略、提高患者安全性并改善预后的关键挑战。机器学习（ML）预测模型在此方面展现出巨大的临床应用潜力。通过整合患者的多模态数据，ML模型能够在免疫治疗启动前提供个性化的风险评估，从而指导临床决策，实现精准医疗。

  1. 高风险患者的早期识别：
    机器学习模型能够从复杂的患者数据中学习，识别出与CIP发生发展相关的高风险因素组合。例如，结合患者的年龄、吸烟史、既往肺部疾病（如间质性肺病ILD）史、合并症、肿瘤负荷以及基线CT影像特征等 6162。对于被模型判定为高风险的患者，临床医生可以对其进行更密切的监测，或考虑采取预防性措施。这种早期识别机制有助于在CIP临床症状出现之前就进行干预，避免疾病进展至严重程度。

  2. 个性化风险评估与治疗方案调整：
    传统的风险评估往往基于群体统计学数据，难以对个体患者提供精准的风险预测。机器学习模型则能够根据每位患者的独特数据特征，生成量化的CIP发生概率或风险等级。基于这种个性化风险评估，医生可以：

    • 调整免疫治疗方案：对于具有极高CIP风险的患者，医生可能考虑调整免疫检查点抑制剂（ICI）的剂量，选择风险较低的ICI单药治疗，或者调整联合用药方案。模型可以帮助评估这些方案对特定患者的风险-收益比。
    • 优化治疗顺序：在涉及放疗与免疫治疗联合的场景中，预测模型可以评估患者发生放疗相关性肺炎（RP）和CIP的风险。研究表明，放疗与免疫治疗的联合会显著增加肺损伤的风险 13。通过ML模型，医生可以权衡先进行免疫治疗再放疗，或先放疗再免疫治疗的顺序，以最大程度地降低肺毒性风险。例如，一项研究表明，整合深度学习提取的影像特征、放射组学特征和临床数据的多模态模型，在预测免疫治疗合并放疗后≥2级症状性肺炎（SP）方面取得了0.922的AUC，显著提升了预测准确性 13。
    • 考虑替代治疗方案：对于风险极高且预估收益不大的患者，医生甚至可以考虑非ICI治疗方案，以避免可能致命的肺部毒性。

  3. 强化监测与早期干预：
    对于经模型评估为高风险的患者，即使决定继续免疫治疗，也可以制定更积极的监测计划，例如增加CT随访频率、定期进行肺功能检查或生物标志物检测。一旦出现早期影像学或临床指标变化，即可立即启动干预措施，如早期使用糖皮质激素，从而降低CIP的严重程度。对3-4级类固醇难治性CIP患者，使用额外的免疫调节剂可以显著降低恶化的风险 62。这种前瞻性的管理策略有望显著改善患者预后。

  4. 优化资源配置：
    通过精准识别高风险患者，医疗资源（如影像学检查、肺功能评估、专科医生咨询）可以更有效地分配给真正需要的患者，避免对所有患者进行过度检查和随访，从而优化医疗成本并提高效率。

总之，机器学习模型在治疗前风险分层方面的应用，使得肺癌免疫治疗从“一刀切”的模式向“个体化精准治疗”迈进。它赋能临床医生在治疗前对CIP风险有更清晰的认识，从而做出更明智的治疗决策，最终为患者带来更安全、更有效的治疗体验。

4.2 与其他类型肺炎的鉴别诊断

肺癌患者在接受免疫治疗过程中，除了可能发生免疫相关性肺炎（CIP）外，还常伴有其他类型的肺炎，如放射性肺炎（RP）、感染性肺炎（特别是对免疫功能低下的患者），甚至可能是由其他原因引起的间质性肺病。这些不同类型肺炎的临床表现和影像学特征往往存在重叠，导致鉴别诊断困难，但它们的治疗和预后却截然不同。例如，CIP通常需要停用免疫检查点抑制剂并使用糖皮质激素治疗，而RP的管理侧重于症状控制，感染性肺炎则需抗生素或抗病毒治疗。错误的诊断不仅会延误最佳治疗时机，还可能加重患者病情。在此背景下，机器学习模型在辅助鉴别诊断方面展现出独特的优势。

  1. 鉴别免疫相关性肺炎（CIP）与放射性肺炎（RP）：
    肺癌患者常常需要接受胸部放疗，放疗后的放射性肺炎是常见的并发症。当患者接受免疫治疗后再发生肺部炎症时，鉴别CIP与RP尤为重要。这两种肺炎在CT影像上存在一些典型差异：

    • 放射性肺炎（RP）：通常局限于放射野内，表现为边界锐利的肺部密度增高或纤维化改变。
    • 免疫相关性肺炎（CIP）：影像学表现更为多样，常表现为双侧、多叶受累，边界通常不锐利，模式包括机化性肺炎、非特异性间质性肺炎等 63。
      机器学习模型能够学习并区分这些细微的影像学特征。一项研究构建了基于放射组学的机器学习模型来区分ICI-相关肺炎和RT-相关肺炎，该模型仅依靠7个放射组学特征即可在曲线下面积（AUC）为0.76的情况下区分ICI-肺炎和RT-肺炎，并能识别出RT+ICI联合治疗后肺炎的主要病因 63。另一项多中心研究中，机器学习模型在区分RP和CIP时表现出适度的性能，测试集AUC为0.8 64。虽然未完全超越放射科医生，但作为辅助工具仍具有价值。特别是对于接受过放疗和免疫治疗的患者，诊断的复杂性增加，模型和医生的鉴别能力均有所下降（AUC 0.54 vs 0.6）64，这提示该领域仍需进一步研究。

  2. 鉴别CIP与感染性肺炎（Infective Pneumonitis, IP），包括COVID-19：
    免疫治疗可能导致免疫功能紊乱，使患者易发生感染，因此鉴别CIP与感染性肺炎（特别是COVID-19等病毒性肺炎）至关重要。

    • COVID-19：常表现为双肺磨玻璃影、实变影，多累及肺外周及下叶。
    • 非COVID-19感染性肺炎：影像学表现多样，常伴有临床感染证据（如发热、白细胞升高）。
      机器学习模型在区分这些不同类型肺炎方面表现出优于人类专家的潜力。在一项研究中，机器学习模型在区分RP与COVID-19相关肺炎、CIP与COVID-19相关肺炎以及CIP与非COVID-19感染性肺炎时，表现均优于放射科医生。例如：
    • RP vs COVID-19：模型测试集AUC为0.92，而两位放射科医生分别为0.8和0.8 64。
    • CIP vs COVID-19：模型测试集AUC为0.68，而两位放射科医生分别为0.43和0.4 64。
    • CIP vs 非COVID-19感染性肺炎：模型测试集AUC为0.71，而两位放射科医生分别为0.55和0.63 64。
      这些结果凸显了机器学习模型作为辅助诊断工具的巨大潜力，尤其是在面对临床表现和影像学模式复杂重叠的肺炎类型时。尽管对于CIP和RP的直接鉴别模型表现仍有提升空间，但其在区分这些肺炎与感染性肺炎时展现出的优越性，能够显著提升诊断的准确性和及时性。

  3. 作为辅助诊断工具的潜力：
    机器学习模型并非旨在取代放射科医生，而是作为一种强大的辅助工具。通过分析大量影像数据和临床信息，模型能够识别出人眼难以察觉的模式和量化特征，为医生提供客观的诊断依据和风险提示。特别是在诊断不确定性较高的情况下，如上述多中心研究所示，模型能够帮助肿瘤学家、放射科医生和胸科医生做出更准确的判断 64。此外，AI工具还可以提高诊断效率，减少放射科医生在阅片过程中的疲劳和主观性差异，从而提升整体诊断水平。例如，在肺栓塞（PE）的检测中，AI辅助工具的加入可以将漏诊率从15.6%显著降低到3.8% 65。

总之，机器学习在肺间质病与其他类型肺炎的鉴别诊断中展现出显著的优势和潜力。通过整合多模态数据并利用强大的模式识别能力，这些模型能够提高诊断的准确性，辅助临床医生做出更及时、更恰当的治疗决策，最终改善患者的预后。

4.3 真实世界验证与临床转化

尽管机器学习模型在实验室环境下展现出优异的预测性能，但其真正的临床价值在于能否在多样化的真实世界场景中保持稳定且可靠的表现，并最终实现临床转化。多中心验证研究是评估模型普适性和稳健性的关键步骤。

  1. 多中心验证研究的结果：
    在肺癌免疫治疗后肺间质病预测领域，多中心验证是衡量模型泛化能力的重要指标。研究表明，一些在单中心开发的高性能模型，在外部独立验证队列中仍能保持良好的预测效能。例如，有研究 66 开发的结合放射组学、临床和定性放射学特征的多模态模型，在训练集上的AUC达到了0.903，在包含来自三个独立中心患者的验证集上仍能保持0.826的AUC。另一项预测免疫检查点抑制剂相关性肺炎（IRP）的机器学习模型在内部验证中取得了AUC 0.81的成绩 67。这些结果证明了机器学习模型从实验室到真实世界应用的潜力。

  2. 模型在不同医疗场景中的普适性：

    • 不同地域与人群：医疗实践和患者特征在全球范围内存在差异，例如，亚洲人群和欧美人群在基因背景、生活习惯、疾病谱和治疗依从性上可能有所不同。因此，模型在某一地区（如中国）训练后，能否在另一地区（如欧美）同样表现出色，是评估其普适性的关键。如果模型能够通过跨区域、跨种族的数据验证，将极大地增强其临床推广的信心。目前的证据表明，在部分研究中，模型在不同地域的验证队列中仍能保持可接受的性能，但仍需更多针对性的跨国多中心研究来全面评估。
    • 不同设备与方案：CT扫描设备型号、扫描协议、重建算法的差异，以及免疫治疗方案（不同ICI药物、联合疗法）的多样性，都可能影响模型的泛化能力。高质量的机器学习模型应具备一定的鲁棒性，能够适应这些变异。标准化数据采集协议和图像预处理方法有助于降低这些异质性带来的影响。

  3. 推动临床落地的关键步骤：
    要将机器学习预测模型从研究成果转化为日常临床实践工具，需要克服技术、规范和接受度等多方面的挑战，并采取以下关键步骤：

    • 标准化数据采集与整合：这是实现模型普适性和持续迭代的基础。制定统一的数据采集标准（包括影像参数、临床指标、实验室结果等），建立多中心、大样本、高质量的数据库。尤其需要规范不同医院、不同设备采集的影像数据预处理流程，以减少数据异质性。
    • 简化模型接口与易用性设计：开发的预测模型应具备用户友好的界面，能够无缝集成到现有的临床信息系统（如PACS、HIS、EHR）中。医生应该能够轻松输入患者数据，并直观地获取预测结果和风险报告。在线IRP风险计算器 67 的设计就是一个很好的范例，它将复杂的模型封装起来，提供便捷的临床工具。
    • 加强模型可解释性：深度学习模型常被认为是“黑箱”，其决策过程难以理解。在临床应用中，医生需要了解模型给出预测结果的依据，以便判断其合理性并承担相应的医疗责任。因此，应进一步发展可解释性AI技术（如SHAP值、LIME、CAM等），使模型能够清晰地揭示哪些特征对预测结果贡献最大。
    • 严格的临床验证与监管批准：模型在上市前必须经过严格的多中心前瞻性临床试验验证，证明其在真实世界中的有效性、安全性和临床获益。同时，需要获得监管机构（如FDA、NMPA）的审批，以确保模型的质量和可靠性。
    • 持续的模型更新与维护：随着新的治疗方案出现和临床实践的演变，模型可能出现“模型漂移”现象。因此，需要建立持续监测和更新机制，利用真实世界数据对模型进行定期再训练和优化，以保持其预测性能和时效性。

通过上述关键步骤的实施，机器学习预测模型有望成为肺癌免疫治疗中重要的辅助决策工具，帮助临床医生实现更加精细化的风险管理和个体化的治疗策略，最终改善肺癌患者的长期生存和生活质量。

5. 挑战与未来研究方向

5.1 当前研究的局限性

尽管机器学习在预测肺癌患者免疫治疗后肺间质病方面展现出巨大潜力，但当前研究仍面临诸多挑战和局限性，这些问题直接影响了模型的可靠性、泛化能力和临床转化。

  1. 数据异质性（Data Heterogeneity）：

    • CT扫描参数差异：不同医疗机构使用的CT扫描设备型号、扫描协议、重建算法及参数设置存在显著差异。这些因素可能导致图像分辨率、噪声水平、对比度等特性各异，从而使得基于影像特征（如放射组学特征、深度学习特征）的模型在不同数据集上表现不稳定。例如，即使是同一患者在不同设备上扫描，其图像特征也可能存在细微差别，这无疑增加了模型泛化的难度。
    • 临床终点定义不统一：肺间质病的诊断标准和分级在不同研究中可能存在差异，例如，对“≥2级肺间质病”的定义可能因研究机构或医生判断标准而异，这导致数据标签的不一致性。此外，对于CIP的鉴别诊断，可能存在与感染性肺炎、放射性肺炎等混淆的情况，这些诊断上的细微偏差都会影响模型训练的准确性。
    • 治疗方案与随访差异：患者接受的免疫治疗方案多样（单药、联合治疗、与放化疗组合），随访时间、随访频率以及不良事件记录的详尽程度也各不相同。这些异质性使得模型难以学习到普适性的治疗响应模式和不良事件发生规律。

  2. 模型可解释性不足（Lack of Model Interpretability）：

    • 深度学习的“黑箱”问题：虽然深度学习模型在处理复杂非结构化数据（如影像）方面表现出色，但其内部决策过程往往不透明，被称为“黑箱模型” 。临床医生需要理解模型做出预测的依据，以信任并采纳其建议。如果模型无法解释其预测结果背后的生物学或临床逻辑，医生在面对模型给出的高风险警示时，可能难以判断其合理性，从而阻碍模型的临床应用。例如，一个模型预测患者有高风险发生CIP，但无法指出是哪些特定的影像纹理特征或临床指标导致了这一判断，医生就很难基于此建议调整治疗方案。虽然SHAP值和注意力机制等可解释性AI技术正在发展，但在实际应用中仍有待提升 68。
    • 特征选择与权重难以量化：传统机器学习模型（如线性回归、决策树）相对容易解释，可以通过特征系数或决策路径来理解。但对于复杂的集成模型或深度学习模型，即使通过特征重要性排名，也难以清晰地揭示特征之间的复杂交互作用，以及它们如何共同驱动最终预测。

  3. 多中心验证样本量有限与外部验证不足：

    • 样本量限制：高质量、大样本的多中心数据集对于训练和验证具有强大泛化能力的机器学习模型至关重要。然而，受限于数据共享、隐私保护和数据采集成本等因素，许多研究往往基于单一中心或数量有限的几个中心的数据集进行开发和验证。小样本量容易导致模型过拟合，使其在训练集上表现良好，但在未见过的新数据上性能显著下降。
    • 外部验证不足：即使是多中心研究，也常常局限于几个特定机构，难以代表真实世界的复杂性和多样性。缺乏独立的、大规模的外部验证（尤其是在不同地域、不同人种群体中的验证），使得模型的普适性和跨平台适用性受到质疑 69。例如，一个在中国患者群体中表现优异的模型，在欧美患者群体中可能因人种差异、基因背景、疾病表型或治疗习惯的不同而效果大打折扣。
    • 研究间的可比性差：由于数据异质性、模型架构差异和评估指标选择的多样性，不同研究之间模型的性能往往难以直接比较，限制了对最佳实践的总结和推广。

这些局限性共同构成了机器学习模型在肺间质病预测领域从研究到临床转化的主要障碍。未来的研究需要着力解决这些问题，以提升模型的稳健性、可靠性和临床实用性。

5.2 未来发展方向

面对当前研究的局限性，未来机器学习在肺癌免疫治疗后肺间质病预测领域的发展将聚焦于提升模型的动态性、可解释性以及多模态数据整合的深度和广度，以期更好地服务于临床实践。

  1. 动态预测模型与实时数据整合：
    目前的预测模型大多基于治疗前的基线数据进行预测，难以反映治疗过程中患者生理状态和疾病进展的动态变化。未来的研究方向之一是开发动态预测模型，即结合治疗中实时更新的影像学（如随访CT）、组学（如循环肿瘤DNA、外周血细胞基因表达）和临床数据。

    • 动态影像数据：通过分析治疗周期内的系列CT影像，捕捉肺部早期微小变化，例如炎症指标的变化、肺部放射组学特征的演变。
    • 动态组学数据：外周血中的免疫细胞亚群、炎症因子（如细胞因子水平 70）、免疫检查点分子表达、循环游离DNA (cfDNA) 或循环肿瘤DNA (ctDNA) 的变化，都可能作为CIP发生的早期生物标志物 71。将这些动态变化的组学数据实时输入模型，可以实现更早期的预警和更精准的风险调整。例如，监测治疗过程中免疫细胞比例或细胞因子水平的异常升高，可能预示着免疫系统过度激活导致的肺部炎症。
    • 优势：动态预测模型能够提供更精细的个体化风险评估，指导临床医生在治疗过程中及时调整方案，实现真正的“预适应”管理，而非被动应对。

  2. 可解释性增强技术（Explainable AI, XAI）：
    为解决深度学习“黑箱”问题，可解释性AI技术将是未来研究的重点。

    • SHapley Additive exPlanations (SHAP) 值：SHAP值通过计算每个特征对模型预测结果的贡献度，为单个预测提供一个统一的、基于博弈论的解释。它可以清晰地展示哪些特征（如特定的CT纹理、某个临床指标或基因表达水平）对模型预测患者将发生CIP起到了正向或负向作用，以及贡献的大小。
    • 注意力机制（Attention Mechanisms）：在深度学习模型中引入注意力机制，可以使模型在处理影像或多模态数据时，将“注意力”集中在对预测结果更重要的区域或特征上。例如，在CT图像中，模型可能会自动突出显示与CIP发生高度相关的肺部区域或病变模式，从而提供视觉上的解释。
    • 概念白盒模型：探索结合深度学习的特征提取能力和传统白盒模型（如决策树、逻辑回归）可解释性的混合模型，例如，通过深度学习提取高级特征后，再输入到易于解释的浅层模型中进行决策。
    • 优势：增强模型可解释性能够建立临床医生对AI工具的信任，帮助他们理解预测依据，将AI建议与临床经验相结合，从而更好地采纳和应用AI结果。

  3. 多组学深度整合与单细胞分析：
    未来的研究将超越传统的单模态或简单多模态融合，实现更深层次的组学数据整合。

    • 多组学深度融合：结合基因组学、转录组学、蛋白质组学和代谢组学等多层次生物信息，通过先进的机器学习方法（如多核学习、图神经网络）挖掘不同组学数据之间的内在关联和相互作用。例如，结合基因变异（基因组）、基因表达（转录组）以及其蛋白质产物（蛋白质组）的数据，可能揭示肺间质病发生更全面的分子机制。
    • 单细胞RNA测序与影像数据结合：单细胞RNA测序技术可以解析肺组织或外周血中单个细胞的基因表达谱，从而识别与CIP发生相关的特定细胞类型（如异常激活的巨噬细胞、T细胞亚群）及其分子特征。将这些单细胞层面的精细信息与宏观的影像学特征相结合，有望建立从微观病理生理机制到宏观影像表现的桥梁，更精准地识别疾病早期风险。例如，通过单细胞分析发现某个患者体内存在高比例的促纤维化巨噬细胞亚群，结合其CT影像中的早期磨玻璃影，可能预示着更高的CIP风险。
    • 优势：深度多组学整合和单细胞分析能够从分子和细胞层面揭示CIP的复杂发病机制，发现更特异、更敏感的生物标志物，为预测和早期干预提供新的靶点和策略。

  4. 真实世界数据驱动的模型迭代与联邦学习：
    模型的持续优化和验证离不开真实世界数据（Real-World Data, RWD）的反馈。

    • 持续模型迭代：建立完善的临床数据反馈机制，利用大规模RWD对已部署的模型进行持续监测、再训练和优化，以应对“模型漂移”现象。这意味着模型不仅仅是静态的预测工具，而是能够随着临床实践的演进而不断学习和改进的动态系统。
    • 联邦学习（Federated Learning）：为解决数据隐私和数据孤岛问题，联邦学习允许多个医疗机构在不共享原始数据的前提下，共同训练一个全局模型。每个机构在本地训练模型，然后只上传模型参数或更新梯度到中央服务器进行聚合。
    • 优势：RWD驱动的迭代和联邦学习有助于构建更具普适性、更稳健且符合隐私保护要求的高质量预测模型，加速其在多中心、多地域临床环境下的推广和应用，最终实现临床转化，为肺癌患者的免疫治疗保驾护航。
内容由 AI 生成，仅供参考，请仔细甄别

参考文献

1First-Line Immunotherapy for Non-Small-Cell Lung Cancer.PubMed

Martin Reck, Jordi Remon, Matthew D Hellmann
J Clin Oncol. 2022 Feb 20;40(6):586-597. doi: 10.1200/JCO.21.01497. Epub 2022 Jan 5.
For patients with metastatic non-small-cell lung cancer (mNSCLC), the last decade has been characterized by critical progress that has contributed to substantially improved survival. In particular, the development of specific antibodies against the programmed death (PD-1) receptor, programmed death-ligand 1 (PD-L1), and the cytotoxic T-lymphocyte-associated protein 4 receptor in the therapeutic strategy of mNSCLC either in first- or in second-line settings have led to unprecedented prolonged survival for a proportion of these patients. Although clinical development of immune checkpoint inhibitors with anti-PD-1 and PD-L1 therapies largely began as monotherapy in the second-line setting, the more recent progress has shifted toward combination approaches in first-line settings as well as the integration of immunotherapy into the clinical paradigm in earlier stages. Today, with the exception of mNSCLC harboring targetable oncogenes, nearly all patients with mNSCLC receive PD-1 or PD-L1 therapy in first-line settings. Here we report the current status of first-line immunotherapy in mNSCLC together with current challenges in selecting the best immunotherapeutic approach for the individual patient.

2Immune-related pulmonary toxicities of checkpoint inhibitors in non-small cell lung cancer: Diagnosis, mechanism, and treatment strategies.PubMed

Xinyu Guo, Shi Chen, Xueyan Wang, et al.
Front Immunol. 2023 Apr 4;14:1138483. doi: 10.3389/fimmu.2023.1138483. eCollection 2023.
Immune checkpoint inhibitors (ICI) therapy based on programmed cell death-1 (PD-1) and programmed cell death ligand 1 (PD-L1) has changed the treatment paradigm of advanced non-small cell lung cancer (NSCLC) and improved the survival expectancy of patients. However, it also leads to immune-related adverse events (iRAEs), which result in multiple organ damage. Among them, the most common one with the highest mortality in NSCLC patients treated with ICI is checkpoint inhibitor pneumonitis (CIP). The respiratory signs of CIP are highly coincident and overlap with those in primary lung cancer, which causes difficulties in detecting, diagnosing, managing, and treating. In clinical management, patients with serious CIP should receive immunosuppressive treatment and even discontinue immunotherapy, which impairs the clinical benefits of ICIs and potentially results in tumor recrudesce. Therefore, accurate diagnosis, detailedly dissecting the pathogenesis, and developing reasonable treatment strategies for CIP are essential to prolong patient survival and expand the application of ICI. Herein, we first summarized the diagnosis strategies of CIP in NSCLC, including the classical radiology examination and the rising serological test, pathology test, and artificial intelligence aids. Then, we dissected the potential pathogenic mechanisms of CIP, including disordered T cell subsets, the increase of autoantibodies, cross-antigens reactivity, and the potential role of other immune cells. Moreover, we explored therapeutic approaches beyond first-line steroid therapy and future direction based on targeted signaling pathways. Finally, we discussed the current impediments, future trends, and challenges in fighting ICI-related pneumonitis.

3Checkpoint Inhibitor Pneumonitis Induced by Anti-PD-1/PD-L1 Therapy in Non-Small-Cell Lung Cancer: Occurrence and Mechanism.PubMed

Jianqiong Yin, Yuanjun Wu, Xue Yang, et al.
Front Immunol. 2022 Apr 7;13:830631. doi: 10.3389/fimmu.2022.830631. eCollection 2022.
Immune checkpointty inhibitors (ICIs), particularly those targeting programmed death 1 (PD-1) and anti-programmed death ligand 1 (PD-L1), enhance the antitumor effect by restoring the function of the inhibited effector T cells and produce durable responses in a large variety of metastatic and late patients with non-small-cell lung cancer. Although often well tolerated, the activation of the immune system results in side effects known as immune-related adverse events (irAEs), which can affect multiple organ systems, including the lungs. The occurrence of severe pulmonary irAEs, especially checkpoint inhibitor pneumonitis (CIP), is rare but has extremely high mortality and often overlaps with the respiratory symptoms and imaging of primary tumors. The development of CIP may be accompanied by radiation pneumonia and infectious pneumonia, leading to the simultaneous occurrence of a mixture of several types of inflammation in the lungs. However, there is a lack of authoritative diagnosis, grading criteria and clarified mechanisms of CIP. In this article, we review the incidence and median time to onset of CIP in patients with non-small-cell lung cancer treated with PD-1/PD-L1 blockade in clinical studies. We also summarize the clinical features, potential mechanisms, management and predictive biomarkers of CIP caused by PD-1/PD-L1 blockade in non-small-cell lung cancer treatment.

4Pneumonitis in Patients Treated With Anti-Programmed Death-1/Programmed Death Ligand 1 Therapy.PubMed

Jarushka Naidoo, Xuan Wang, Kaitlin M Woo, et al.
J Clin Oncol. 2017 Mar;35(7):709-717. doi: 10.1200/JCO.2016.68.2005. Epub 2016 Sep 30.
Purpose Pneumonitis is an uncommon but potentially fatal toxicity of anti-programmed death-1 (PD-1)/programmed death ligand 1 (PD-L1) monoclonal antibodies (mAbs). Clinical, radiologic, and pathologic features are poorly described. Methods Patients who received anti-PD-1/PD-L1 monotherapy or in combination with anti-cytotoxic T-cell lymphocyte-4 mAb were identified at two institutions (Memorial Sloan Kettering Cancer Center: advanced solid cancers, 2009 to 2014, and Melanoma Institute of Australia: melanomas only, 2013 to 2015). Pneumonitis was diagnosed by the treating investigator; cases with confirmed malignant lung infiltration or infection were excluded. Clinical, radiologic, and pathologic features of pneumonitis were collected. Associations among pneumonitis incidence, therapy received, and underlying malignancy were examined with Fisher's exact test as were associations between pneumonitis features and outcomes. Results Of 915 patients who received anti-PD-1/PD-L1 mAbs, pneumonitis developed in 43 (5%; 95% CI, 3% to 6%; Memorial Sloan Kettering Cancer Center, 27 of 578 [5%]; Melanoma Institute of Australia, 16 of 337 [5%]). Time to onset of pneumonitis ranged from 9 days to 19.2 months. The incidence of pneumonitis was higher with combination immunotherapy versus monotherapy (19 of 199 [10%] v 24 of 716 [3%]; P < .01). Incidence was similar in patients with melanoma and non-small-cell lung cancer (overall, 26 of 532 [5%] v nine of 209 [4%]; monotherapy, 15 of 417 v five of 152 [ P = 1.0]; combination, 11 of 115 v four of 57 [ P = .78]). Seventy-two percent (31 of 43) of cases were grade 1 to 2, and 86% (37 of 43) improved/resolved with drug holding/immunosuppression. Five patients worsened clinically and died during the course of pneumonitis treatment; proximal cause of death was pneumonitis (n = 1), infection related to immunosuppression (n = 3), or progressive cancer (n = 1). Radiologic and pathologic features of pneumonitis were diverse. Conclusion Pneumonitis associated with anti-PD-1/PD-L1 mAbs is a toxicity of variable onset and clinical, radiologic, and pathologic appearances. It is more common when anti-PD-1/PD-L1 mAbs are combined with anti-cytotoxic T-cell lymphocyte-4 mAb. Most events are low grade and improve/resolve with drug holding/immunosuppression. Rarely, pneumonitis worsens despite immunosuppression, and may result in infection and/or death.

5[Pneumonitis Induced by Immune Checkpoint Inhibitors].PubMed

Eriko Miyawaki, Hirotsugu Kenmotsu
Gan To Kagaku Ryoho. 2018 Jul;45(7):1021-1026.
Immune checkpoint inhibitors(ICIs)have shown efficacy in many clinical trials and have an important role in the treatment for advanced cancers.Many physicians recently have chances to use ICIs across tumor types.ICIs are associated with a unique set of toxic effects, which are defined as immune-related adverse events(irAEs).Drug -induced pneumonitis is considered to be one of important irAEs.Pneumonitis as irAE has been not so frequently observed, and all Grade in 0~10%and Grade 3 or 4 in 0~3% of patients receiving ICIs.However severe or life-threatening pneumonitis was observed in some cases, so that pneumonitis caused by ICIs has been clinically important.The incidence of pneumonitis caused by ICIs for Japanese patients is similar to previous reports.Symptoms of pneumonitis by ICIs are considered to be dyspnea and cough.Some patients are diagnosed as asymptomatic pneumonitis with abnormal radiographic findings.Cryptogenic organizing pneumonia(COP) pattern has been reported to be most frequently observed in radiographic patterns of pneumonitis caused by ICIs(43~65%). Since some cases shows pneumonitis with different radiographic pattern from other drug-induced pneumonitis, blood test and bronchoscopy may be useful for diagnosis.It has been recommended to treat patients with ICIs-related pneumonitis according to severity of pneumonitis.In most patients, pneumonitis is improved with corticosteroids and withdrawal of ICIs, though a few patients with severe pneumonitis may need immunosuppressants.It has been recommended for medical oncologists to consult and cooperate with pulmonologists for diagnosis and treatment of ICIs-related interstitial lung disease(ILD). The mechanism of pneumonitis induced by ICIs may differ from those of other drug-induced pneumonitis.Therefore, medical oncologists should be careful for ICIs-related pneumonitis as a new disease category in the treatment of ICIs.

6Immune Checkpoint Inhibitor Therapy-related Pneumonitis: Patterns and Management.PubMed

Kevin R Kalisz, Nikhil H Ramaiya, Kai R Laukamp, et al.
Radiographics. 2019 Nov-Dec;39(7):1923-1937. doi: 10.1148/rg.2019190036. Epub 2019 Oct 4.
In recent years, the use of immune checkpoint inhibitor (ICI) therapy has rapidly grown, with increasing U.S. Food and Drug Administration approvals of a variety of agents used as first- and second-line treatments of various malignancies. ICIs act through a unique mechanism of action when compared with those of conventional chemotherapeutic agents. ICIs target the cell surface receptors cytotoxic T-lymphocyte antigen-4, programmed cell death protein 1, or programmed cell death ligand 1, which result in immune system-mediated destruction of tumor cells. Immune-related adverse events are an increasingly recognized set of complications of ICI therapy that may affect any organ system. ICI therapy-related pneumonitis is an uncommon but important complication of ICI therapy, with potential for significant morbidity and mortality. As the clinical manifestation is often nonspecific, CT plays an important role in diagnosis and triage. Several distinct radiographic patterns of pneumonitis have been observed: organizing pneumonia, nonspecific interstitial pneumonia, hypersensitivity pneumonitis, acute interstitial pneumonia-acute respiratory distress syndrome, bronchiolitis, and radiation recall pneumonitis. Published guidelines outline the treatment of ICI therapy-related pneumonitis based on the severity of symptoms. Treatment is often effective, although recurrence is possible. This article reviews the mechanism of ICIs and ICI therapy complications, with subsequent management techniques and illustrations of the various radiologic patterns of ICI-therapy related pneumonitis.RSNA, 2019.

7Immune checkpoint inhibitor-related interstitial lung disease in patients with advanced non-small cell lung cancer: systematic review of characteristics, incidence, risk factors, and management.PubMed

Seohyun Kim, Jeong Uk Lim
J Thorac Dis. 2022 May;14(5):1684-1695. doi: 10.21037/jtd-22-93.
BACKGROUND: Immune checkpoint inhibitors (ICIs) are widely used in cancers with or without other treatments. Immune-related adverse events (irAEs) have been reported as side effects of ICIs and involve various organs. Pneumonitis, which is also called ICI-related interstitial lung disease (ILD), is one of the life-threatening adverse events. In this report, we reviewed the safety of ICIs and risk factors for ICI-related ILD in non-small cell lung cancer (NSCLC) patients. METHODS: Databases (The National Center for Biotechnology Information, PubMed, Cochrane Library, Google Scholar, and Embase) were searched for the literature on pulmonary adverse events and immunotherapy following PRISMA guidelines. All studies published in English between January 2016 and June 2021 were included. RESULTS: One-hundred twenty-five articles were included at final. Pre-existing ILD as well as asthma and chronic obstructive lung disease (COPD) were associated with high risk for the development of ICI-related ILD, however, it did not affect the prognosis. The treatment of ICI-related ILD is different according to the severity grades, which includes discontinuation of ICI, corticosteroids, and for steroid-refractory ILD, other immunosuppressants. Rechallenging of ICI should be carefully considered in selected patients. DISCUSSION: Patients with pre-existing lung disease and poor lung function need more attention for the development of ICI-related ILD. It is necessary to be aware of clinical manifestation of ICI-related ILD with prompt management.

8Risk Factors for Immune Checkpoint Inhibitor-Related Pneumonitis in Cancer Patients: A Systemic Review and Meta-Analysis.PubMed

Peining Zhou, Xiang Zhao, Guangfa Wang
Respiration. 2022;101(11):1035-1050. doi: 10.1159/000526141. Epub 2022 Sep 15.
BACKGROUND: Immune checkpoint inhibitors (ICIs) can lead to one of the common and quite serious immune-related adverse events (irAEs) in a real-world setting, namely, checkpoint inhibitor pneumonitis (CIP). OBJECTIVE: We aimed to investigate the potential risk factors for CIP in cancer patients treated by ICIs and quantify the association. METHODS: We conducted a systematic literature review in PubMed, EMBASE, and Web of Science from January 1, 2000, to March 20, 2022, with no study design restrictions. Studies evaluating the risk factors for CIP in cancer patients treated with ICIs-containing regimes were included. Odds ratios (ORs) with 95% confidence intervals (CIs) were calculated for risk factors of CIP by using random-effect model. Heterogeneity was assessed by sensitivity analysis and subgroup analysis regarding CIP severity, geographical regions, cancer types, treatment regimes, and ICI types. RESULTS: A total of 35 studies comprising 142,703 patients were eventually included. The incidence of grade I-V and grade III-V CIP in cancer patients was 0.16 (95% CI: 0.14-0.18) and 0.06 (95% CI: 0.05-0.08), respectively. When combining the adjusted ORs, the following risk factors were significantly associated with the development of CIP: squamous cell carcinoma (OR: 1.31, 95% CI: 1.18-1.45), previous thoracic radiotherapy (OR: 2.07, 95% CI: 1.34-3.19), preexisting radiation-induced pneumonitis (OR: 3.62, 95% CI: 1.53-8.58), preexisting respiratory disease (OR: 2.43, 95% CI: 1.45-4.07), preexisting interstitial lung disease (OR: 5.78, 95% CI: 3.08-10.85), preexisting ground glass attenuation (OR: 11.48, 95% CI: 1.13-116.74), preexisting honeycombing (OR: 6.11, 95% CI: 2.37-15.79), preexisting pulmonary emphysema (OR: 2.72, 95% CI: 1.00-7.36), use of pembrolizumab (OR: 2.89, 95% CI: 1.56-5.35, I2 = 0%), high PD-L1 expression (OR: 3.59, 95% CI: 1.23-10.50), and hypoalbuminemia (OR: 0.3, 95% CI: 0.14-0.64). When including the crude ORs, smoking history (OR: 1.39, 95% CI: 1.14-1.71), neutrophil-lymphocyte ratio (OR: 1.04, 95% CI: 1.01-1.08), and c-reactive protein (OR: 1.08, 95% CI: 1.01-1.16) were also risk factors for CIP. CONCLUSION: Histological characteristics, specific previous lung diseases and treatment history, treatment regimen, PD-L1 expression level, and serological biomarkers are all closely associated with the development of CIP. These findings would be useful for oncologists to optimize the appropriate options of ICIs and close monitoring during immunotherapy treatments, particularly for patients identified as having a higher risk for CIP.

9Programmed cell death 1 (PD-1)/PD-ligand 1(PD-L1) inhibitors-related pneumonitis in patients with advanced non-small cell lung cancer.PubMed

Yuxin Sun, Chi Shao, Shan Li, et al.
Asia Pac J Clin Oncol. 2020 Dec;16(6):299-304. doi: 10.1111/ajco.13380. Epub 2020 Aug 5.
Immune-related pneumonitis is an uncommon but potentially fatal immune-related adverse event in advanced non-small cell lung cancer (NSCLC) patients during treatment with anti-programmed cell death 1 (PD-1) and programmed cell death-ligand 1 (PD-L1). Underlying emphysema, interstitial lung disease (ILD), and previous radiation therapy for lung cancer might be factors precipitating immune-related pneumonitis. The incidence of immune-related pneumonitis is reported to be higher in those treated with PD-1 inhibitors than in those treated with anti-PD-L1 inhibitors. Early detection and diagnosis and appropriate management according to the severity are critical to improving the prognosis. The first-line physicians, including the primary responsible oncologists, family doctors, emergency physicians and NSCLC patients should be trained to identify and report symptoms of immune-related pneumonitis as early as possible. Multidisciplinary treatment teams involving clinicians (including ILD specialists and lung cancer specialists), radiologists and pathologists are recommended for the treatment of immune-related pneumonitis.

10Steroid-refractory PD-(L)1 pneumonitis: incidence, clinical features, treatment, and outcomes.PubMed

Aanika Balaji, Melinda Hsu, Cheng Ting Lin, et al.
J Immunother Cancer. 2021 Jan;9(1). doi: 10.1136/jitc-2020-001731.
BACKGROUND: Immune-checkpoint inhibitor (ICI)-pneumonitis that does not improve or resolve with corticosteroids and requires additional immunosuppression is termed steroid-refractory ICI-pneumonitis. Herein, we report the clinical features, management and outcomes for patients treated with intravenous immunoglobulin (IVIG), infliximab, or the combination of IVIG and infliximab for steroid-refractory ICI-pneumonitis. METHODS: Patients with steroid-refractory ICI-pneumonitis were identified between January 2011 and January 2020 at a tertiary academic center. ICI-pneumonitis was defined as clinical or radiographic lung inflammation without an alternative diagnosis, confirmed by a multidisciplinary team. Steroid-refractory ICI-pneumonitis was defined as lack of clinical improvement after high-dose corticosteroids for 48 hours, necessitating additional immunosuppression. Serial clinical, radiologic (CT imaging), and functional features (level-of-care, oxygen requirement) were collected preadditional and postadditional immunosuppression. RESULTS: Of 65 patients with ICI-pneumonitis, 18.5% (12/65) had steroid-refractory ICI-pneumonitis. Mean age at diagnosis of ICI-pneumonitis was 66.8 years (range: 35-85), 50% patients were male, and the majority had lung carcinoma (75%). Steroid-refractory ICI-pneumonitis occurred after a mean of 5 ICI doses from PD-(L)1 start (range: 3-12 doses). The most common radiologic pattern was diffuse alveolar damage (DAD: 50%, 6/12). After corticosteroid failure, patients were treated with: IVIG (n=7), infliximab (n=2), or combination IVIG and infliximab (n=3); 11/12 (91.7%) required ICU-level care and 8/12 (75%) died of steroid-refractory ICI-pneumonitis or infectious complications (IVIG alone=3/7, 42.9%; infliximab alone=2/2, 100%; IVIG + infliximab=3/3, 100%). All five patients treated with infliximab (5/5; 100%) died from steroid-refractory ICI-pneumonitis or infectious complications. Mechanical ventilation was required in 53% of patients treated with infliximab alone, 80% of those treated with IVIG + infliximab, and 25.5% of those treated with IVIG alone. CONCLUSIONS: Steroid-refractory ICI-pneumonitis constituted 18.5% of referrals for multidisciplinary irAE care. Steroid-refractory ICI-pnuemonitis occurred early in patients' treatment courses, and most commonly exhibited a DAD radiographic pattern. Patients treated with IVIG alone demonstrated an improvement in both level-of-care and oxygenation requirements and had fewer fatalities (43%) from steroid-refractory ICI-pneumonitis when compared to treatment with infliximab (100% mortality).

11AI in Health: State of the Art, Challenges, and Future Directions.PubMed

Fei Wang, Anita Preininger
Yearb Med Inform. 2019 Aug;28(1):16-26. doi: 10.1055/s-0039-1677908. Epub 2019 Aug 16.
INTRODUCTION: Artificial intelligence (AI) technologies continue to attract interest from a broad range of disciplines in recent years, including health. The increase in computer hardware and software applications in medicine, as well as digitization of health-related data together fuel progress in the development and use of AI in medicine. This progress provides new opportunities and challenges, as well as directions for the future of AI in health. OBJECTIVE: The goals of this survey are to review the current state of AI in health, along with opportunities, challenges, and practical implications. This review highlights recent developments over the past five years and directions for the future. METHODS: Publications over the past five years reporting the use of AI in health in clinical and biomedical informatics journals, as well as computer science conferences, were selected according to Google Scholar citations. Publications were then categorized into five different classes, according to the type of data analyzed. RESULTS: The major data types identified were multi-omics, clinical, behavioral, environmental and pharmaceutical research and development (R&D) data. The current state of AI related to each data type is described, followed by associated challenges and practical implications that have emerged over the last several years. Opportunities and future directions based on these advances are discussed. CONCLUSION: Technologies have enabled the development of AI-assisted approaches to healthcare. However, there remain challenges. Work is currently underway to address multi-modal data integration, balancing quantitative algorithm performance and qualitative model interpretability, protection of model security, federated learning, and model bias.

12Predicting cancer risk using machine learning on lifestyle and genetic data.PubMed

Mohamed Abdelmoaty Ahmed, Ahmed AbdelMoety, Asmaa Mohamed Ahmed Soliman
Sci Rep. 2025 Aug 19;15(1):30458. doi: 10.1038/s41598-025-15656-8.
Cancer remains one of the leading causes of mortality worldwide, where early detection significantly improves patient outcomes and reduces treatment burden. This study investigates the application of Machine Learning (ML) techniques to predict cancer risk based on a combination of genetic and lifestyle factors. A structured dataset of 1,200 patient records was used, comprising features such as age, gender, Body Mass Index (BMI), smoking status, alcohol intake, physical activity, genetic risk level, and personal history of cancer. A full end-to-end ML pipeline was implemented, encompassing data exploration, preprocessing, feature scaling, model training, and evaluation using stratified cross-validation and a separate test set. Nine supervised learning algorithms were evaluated and compared, including Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), Support Vector Machines (SVMs), and several ensemble methods. Among these, Categorical Boosting (CatBoost) achieved the highest predictive performance, with a test accuracy of 98.75% and an F1-score of 0.9820, outperforming both traditional and other advanced models. Feature importance analysis confirmed the strong influence of cancer history, genetic risk, and smoking status on prediction outcomes. The findings highlight the effectiveness of boosting-based ensemble models in capturing complex interactions within health data and support their potential use in personalized cancer risk assessment. This research underscores the value of integrating genetic and modifiable lifestyle variables into predictive modeling to enhance early detection and preventive healthcare strategies.

13Multimodal data deep learning method for predicting symptomatic pneumonitis caused by lung cancer radiotherapy combined with immunotherapy.PubMed

Mingyu Yang, Jianli Ma, Chengcheng Zhang, et al.
Front Immunol. 2025 Jan 8;15:1492399. doi: 10.3389/fimmu.2024.1492399. eCollection 2024.
OBJECTIVES: The pairing of immunotherapy and radiotherapy in the treatment of locally advanced nonsmall cell lung cancer (NSCLC) has shown promise. By combining radiotherapy with immunotherapy, the synergistic effects of these modalities not only bolster antitumor efficacy but also exacerbate lung injury. Consequently, developing a model capable of accurately predicting radiotherapy- and immunotherapy-related pneumonitis in lung cancer patients is a pressing need. Depth image features extracted from deep learning, combined with radiomics and clinical characteristics, were used to create a deep learning model. This model was developed to forecast symptomatic pneumonitis (SP) (≥Grade 2) in lung cancer patients undergoing thoracic radiotherapy in combination with immunotherapy. METHODS: The prediction was based on CT scans taken prior to the start of thoracic radiotherapy. Retrospective collection of clinical data was conducted on 261 lung cancer patients undergoing a combination of thoracic radiotherapy and immunotherapy from January 2018 to May 2023. Imaging data in the form of pre-RT-CT scans were obtained for all individuals included in the study. The region of interest (ROI) in the lung parenchyma was outlined separately from the tumor volume, and standard radiomic features were obtained through the use of 3D Slicer software. In addition, the images were cropped to a uniform size of 224x224 pixels. Data augmentation techniques, including random horizontal flipping, were employed. The normalized image data was then input into a pre-trained deep residual network, ResNet34, which utilized convolutional layers and global average pooling layers for deep feature extraction. A five-fold cross-validation approach was implemented to construct the model, automatically splitting the dataset into training and validation sets at an 8:2 ratio. This process was repeated five times, and the results from these iterations were aggregated to compute the average values of performance metrics, thereby assessing the overall performance and stability of the model. RESULTS: The multimodal fusion model developed in this research, which incorporated depth image characteristics, radiomics properties, and clinical data, demonstrated an AUC of 0.922 (95% CI: 0.902-0.945, P value < 0.001). This amalgamated model surpassed the performance of the radiomic feature model (AUC 0.811, 95% CI: 0.786-0.832, P value < 0.001), the clinical information model (AUC 0.711, 95% CI: 0.682-0.753, P value < 0.001), as well as the model that integrated omics attributes with clinical data (AUC 0.872, 95% CI: 0.845-0.896, P value < 0.001) utilizing deep neural networks (DNNs). Comparatively, the radiomic feature model based on random forest (RF) yielded an AUC of 0.576, with a 95% confidence interval of 0.523-0.628. The clinical information model based on RF had an AUC of 0.525, with a 95% confidence interval of 0.479-0.572. When both radiomic features and clinical information were combined in a model based on RF, the AUC improved slightly to 0.611, with a 95% confidence interval of 0.566-0.652. CONCLUSIONS: In this study, a deep neural network-based multimodal fusion model improved the prediction performance compared to traditional radiomics. The model accurately predicted Grade 2 or higher SP in lung cancer patients undergoing radiotherapy combined with immunotherapy.

14From Silos to Synthesis: A comprehensive review of domain adaptation strategies for multi-source data integration in healthcare.PubMed

Shelia Rahman Tuly, Sima Ranjbari, Ekrem Alper Murat, et al.
Comput Biol Med. 2025 Jun;191:110108. doi: 10.1016/j.compbiomed.2025.110108. Epub 2025 Apr 9.
BACKGROUND: The integration of data from diverse sources is not only crucial for addressing data scarcity in health informatics but also enables the use of complementary information from multiple datasets. However, the isolated nature of data collected from disparate sources (referred to as 'Silos') presents significant challenges in multi-source data integration due to inherent heterogeneity and differences in data structures, formats, and standards. Domain adaptation emerges as a key framework to transition from 'Silos' to 'Synthesis' by measuring and mitigating such discrepancies, enabling uniform representation and harmonization of multi-source data. METHODS: This study explores different approaches to healthcare data integration, highlighting the challenges associated with each type and discussing both general-purpose and healthcare-specific adaptation methods. We examine key research challenges and evaluate leading domain adaptation approaches, demonstrating their effectiveness and limitations in advancing healthcare data integration. RESULTS: The findings highlight the potential of domain adaptation methods to significantly improve healthcare data integration while laying a foundation for future research. CONCLUSION: Current research often lacks a comprehensive analysis of how domain adaptation can effectively address the challenges associated with integrating multi-source and multi-modal healthcare datasets. This study serves as a valuable resource for healthcare professionals and researchers, providing guidance on leveraging domain adaptation techniques to mitigate domain discrepancies in healthcare data integration.

15The use of machine learning and nonlinear statistical tools for ADME prediction.PubMed

Yojiro Sakiyama
Expert Opin Drug Metab Toxicol. 2009 Feb;5(2):149-69. doi: 10.1517/17425250902753261.
Absorption, distribution, metabolism and excretion (ADME)-related failure of drug candidates is a major issue for the pharmaceutical industry today. Prediction of ADME by in silico tools has now become an inevitable paradigm to reduce cost and enhance efficiency in pharmaceutical research. Recently, machine learning as well as nonlinear statistical tools has been widely applied to predict routine ADME end points. To achieve accurate and reliable predictions, it would be a prerequisite to understand the concepts, mechanisms and limitations of these tools. Here, we have devised a small synthetic nonlinear data set to help understand the mechanism of machine learning by 2D-visualisation. We applied six new machine learning methods to four different data sets. The methods include Naive Bayes classifier, classification and regression tree, random forest, Gaussian process, support vector machine and k nearest neighbour. The results demonstrated that ensemble learning and kernel machine displayed greater accuracy of prediction than classical methods irrespective of the data set size. The importance of interaction with the engineering field is also addressed. The results described here provide insights into the mechanism of machine learning, which will enable appropriate usage in the future.

16Hash Transformation and Machine Learning-Based Decision-Making Classifier Improved the Accuracy Rate of Automated Parkinson's Disease Screening.PubMed

Tsung-Lung Yang, Chia-Hung Lin, Wei-Ling Chen, et al.
IEEE Trans Neural Syst Rehabil Eng. 2020 Jan;28(1):72-82. doi: 10.1109/TNSRE.2019.2950143. Epub 2019 Oct 29.
Digitalized hand-drawn pattern is a noninvasive and reproducible assistive manner to obtain hand actions and motions for evaluating functional tremors and upper-limb movement disorders. In this study, spirals and straight lines in polar coordinates are used to extract polar expression features such as the key parameters deviation (cm) and accumulation angle (rad). These parameters are quantitative manner to scale the variations of functional tremors in normal control subjects and patients with Parkinson's disease (PD) and essential tremor (ET). However, difficulty arises in using nonlinear polar expression features in the two-dimensional feature space to separate normal control subjects from those with PD and ET. To solve the nonlinear separable classification problem, hash transformation is used to map polar expression features to a high-dimensional space using hash weighing function and modulo operation. Then, a machine learning method, such as the generalized regression neural network (GRNN), is implemented to train a decision-making classifier using the particle swarm optimization (PSO) algorithm for possible class assessment. With the enrolled data from 50 subjects, the fivefold cross validation, mean true positive, mean true negative, and mean hit rates of 98.93%, 98.96%, and 98.93%, respectively, are obtained to quantify the performance of the proposed decision-making classifier to identify normal controls and subjects with PD or ET. The experimental results indicate that the proposed screening model can improve the accuracy rate compared with the conventional machine learning classifier.

17Long-Term Predictive Modelling of the Craniofacial Complex Using Machine Learning on 2D Cephalometric Radiographs.PubMed

Michael Myers, Michael D Brown, Sarkhan Badirli, et al.
Int Dent J. 2025 Feb;75(1):236-247. doi: 10.1016/j.identj.2024.12.023. Epub 2025 Jan 5.
OBJECTIVE: This study aimed to predict long-term growth-related changes in skeletal and dental relationships within the craniofacial complex using machine learning (ML) models. MATERIALS AND METHODS: Cephalometric radiographs from 301 subjects, taken at pre-pubertal (T1, age 11) and post-pubertal stages (T2, age 18), were analysed. Three ML models-Lasso regression, Random Forest, and Support Vector Regression (SVR)-were trained on a subset of 240 subjects, while 61 subjects were used for testing. Model performance was evaluated using mean absolute error (MAE), intraclass correlation coefficients (ICCs), and clinical thresholds (2 mm or 2°). RESULTS: MAEs for skeletal measurements ranged from 1.36° (maxilla to cranial base angle) to 4.12 mm (mandibular length), and for dental measurements from 1.26 mm (lower incisor position) to 5.40° (upper incisor inclination). ICCs indicated moderate to excellent agreement between actual and predicted values. The highest prediction accuracy within the 2 mm or 2° clinical thresholds was achieved for maxilla to cranial base angle (80%), lower incisor position (75%), and maxilla to mandible angle (70%). Pre-pubertal measurements and sex consistently emerged as the most important predictive factors. CONCLUSIONS: ML models demonstrated the ability to predict post-pubertal values for maxilla to cranial base, mandible to cranial base, maxilla to mandible angles, upper and lower incisor positions, and upper face height with a clinically acceptable margin of 2 mm or 2°. Prediction accuracy was higher for skeletal relationships compared to dental relationships over the 8-year growth period. Pre-pubertal values of the measurements and sex emerged consistently as the most important predictors of the post-pubertal values.

18Machine Learning in Hypertrophic Cardiomyopathy: Nonlinear Model From Clinical and CMR Features Predicting Cardiovascular Events.PubMed

Kankan Zhao, Yanjie Zhu, Xiuyu Chen, et al.
JACC Cardiovasc Imaging. 2024 Aug;17(8):880-893. doi: 10.1016/j.jcmg.2024.04.013. Epub 2024 Jul 10.
BACKGROUND: The cumulative burden of hypertrophic cardiomyopathy (HCM) is significant, with a noteworthy percentage (10%-15%) of patients with HCM per year experiencing major adverse cardiovascular events (MACEs). A current risk stratification scheme for HCM had only limited accuracy in predicting sudden cardiac death (SCD) and failed to account for a broader spectrum of adverse cardiovascular events and cardiac magnetic resonance (CMR) parameters. OBJECTIVES: This study sought to develop and evaluate a machine learning (ML) framework that integrates CMR imaging and clinical characteristics to predict MACEs in patients with HCM. METHODS: A total of 758 patients with HCM (67% male; age 49 ± 14 years) who were admitted between 2010 and 2017 from 4 medical centers were included. The ML model was built on the internal discovery cohort (533 patients with HCM, admitted to Fuwai Hospital, Beijing, China) by using the light gradient-boosting machine and internally evaluated using cross-validation. The external test cohort consisted of 225 patients with HCM from 3 medical centers. A total of 14 CMR imaging features (strain and late gadolinium enhancement [LGE]) and 23 clinical variables were evaluated and used to inform the ML model. MACEs included a composite of arrhythmic events, SCD, heart failure, and atrial fibrillation-related stroke. RESULTS: MACEs occurred in 191 (25%) patients over a median follow-up period of 109.0 months (Q1-Q3: 73.0-118.8 months). Our ML model achieved areas under the curve (AUCs) of 0.830 and 0.812 (internally and externally, respectively). The model outperformed the classic HCM Risk-SCD model, with significant improvement (P < 0.001) of 22.7% in the AUC. Using the cubic spline analysis, the study showed that the extent of LGE and the impairment of global radial strain (GRS) and global circumferential strain (GCS) were nonlinearly correlated with MACEs: an elevated risk of adverse cardiovascular events was observed when these parameters reached the high enough second tertiles (11.6% for LGE, 25.8% for GRS, -17.3% for GCS). CONCLUSIONS: ML-empowered risk stratification using CMR and clinical features enabled accurate MACE prediction beyond the classic HCM Risk-SCD model. In addition, the nonlinear correlation between CMR features (LGE and left ventricular pressure gradient) and MACEs uncovered in this study provides valuable insights for the clinical assessment and management of HCM.

19A machine learning-based prediction model for sepsis-associated delirium in intensive care unit patients with sepsis-associated acute kidney injury.PubMed

Shuangjiang Yu, Xuming Pan, Manyuan Zhang, et al.
Ren Fail. 2025 Dec;47(1):2514186. doi: 10.1080/0886022X.2025.2514186. Epub 2025 Jun 2.
Sepsis-associated acute kidney injury (SA-AKI) patients in the ICU often suffer from sepsis-associated delirium (SAD), which is linked to unfavorable outcomes. This research aimed to develop a machine learning-based model for early SAD prediction in SA-AKI patients. Data was sourced from the Medical Information Mart for Intensive Care IV (MIMIC-IV) and eICU Collaborative Research Database (eICU-CRD). Various models, including logistic regression, extreme gradient boosting (XGBoost), random forest, k-nearest neighbors, support vector machine, decision tree, and naive Bayes, were constructed and evaluated. The XGBoost model emerged as the best, with an internal validation AUROC of 0.775 and an external validation AUROC of 0.687. Unlike traditional delirium assessments, this model enables earlier SAD prediction and is suitable for patients who are hard to assess conventionally.

20Analysis of Machine Learning Techniques for Heart Failure Readmissions.PubMed

Bobak J Mortazavi, Nicholas S Downing, Emily M Bucholz, et al.
Circ Cardiovasc Qual Outcomes. 2016 Nov;9(6):629-640. doi: 10.1161/CIRCOUTCOMES.116.003039. Epub 2016 Nov 8.
BACKGROUND: The current ability to predict readmissions in patients with heart failure is modest at best. It is unclear whether machine learning techniques that address higher dimensional, nonlinear relationships among variables would enhance prediction. We sought to compare the effectiveness of several machine learning algorithms for predicting readmissions. METHODS AND RESULTS: Using data from the Telemonitoring to Improve Heart Failure Outcomes trial, we compared the effectiveness of random forests, boosting, random forests combined hierarchically with support vector machines or logistic regression (LR), and Poisson regression against traditional LR to predict 30- and 180-day all-cause readmissions and readmissions because of heart failure. We randomly selected 50% of patients for a derivation set, and a validation set comprised the remaining patients, validated using 100 bootstrapped iterations. We compared C statistics for discrimination and distributions of observed outcomes in risk deciles for predictive range. In 30-day all-cause readmission prediction, the best performing machine learning model, random forests, provided a 17.8% improvement over LR (mean C statistics, 0.628 and 0.533, respectively). For readmissions because of heart failure, boosting improved the C statistic by 24.9% over LR (mean C statistic 0.678 and 0.543, respectively). For 30-day all-cause readmission, the observed readmission rates in the lowest and highest deciles of predicted risk with random forests (7.8% and 26.2%, respectively) showed a much wider separation than LR (14.2% and 16.4%, respectively). CONCLUSIONS: Machine learning methods improved the prediction of readmission after hospitalization for heart failure compared with LR and provided the greatest predictive range in observed readmission rates.

21Advancing non-alcoholic fatty liver disease prediction: a comprehensive machine learning approach integrating SHAP interpretability and multi-cohort validation.PubMed

Bo Yang, Huaguan Lu, Yinghui Ran
Front Endocrinol (Lausanne). 2024 Oct 8;15:1450317. doi: 10.3389/fendo.2024.1450317. eCollection 2024.
INTRODUCTION: Non-alcoholic fatty liver disease (NAFLD) represents a major global health challenge, often undiagnosed because of suboptimal screening tools. Advances in machine learning (ML) offer potential improvements in predictive diagnostics, leveraging complex clinical datasets. METHODS: We utilized a comprehensive dataset from the Dryad database for model development and training and performed external validation using data from the National Health and Nutrition Examination Survey (NHANES) 2017-2020 cycles. Seven distinct ML models were developed and rigorously evaluated. Additionally, we employed the SHapley Additive exPlanations (SHAP) method to enhance the interpretability of the models, allowing for a detailed understanding of how each variable contributes to predictive outcomes. RESULTS: A total of 14,913 participants were eligible for this study. Among the seven constructed models, the light gradient boosting machine achieved the highest performance, with an area under the receiver operating characteristic curve of 0.90 in the internal validation set and 0.81 in the external NHANES validation cohort. In detailed performance metrics, it maintained an accuracy of 87%, a sensitivity of 92.9%, and an F1 score of 0.92. Key predictive variables identified included alanine aminotransferase, gammaglutamyl transpeptidase, triglyceride glucose-waist circumference, metabolic score for insulin resistance, and HbA1c, which are strongly associated with metabolic dysfunctions integral to NAFLD progression. CONCLUSIONS: The integration of ML with SHAP interpretability provides a robust predictive tool for NAFLD, enhancing the early identification and potential management of the disease. The model's high accuracy and generalizability across diverse populations highlight its clinical utility, though future enhancements should include longitudinal data and lifestyle factors to refine risk assessments further.

22Interpretable Machine Learning Prediction Model for Predicting Mortality Risk of ICU Patients With Pressure Ulcers Based on the Braden Scale: A Clinical Study Based on MIMIC-IV.PubMed

Binyan Chen, Jinghao Zhou, Shengzhang Chen, et al.
J Clin Nurs. 2025 Aug;34(8):3353-3369. doi: 10.1111/jocn.17860. Epub 2025 Jun 12.
AIMS: This study was to create an interpretable machine learning model to predict the risk of mortality within 90 days for ICU patients suffering from pressure ulcers. DESIGN: We retrospectively analysed 1774 ICU pressure ulcer patients from the Medical Information Mart for Intensive Care (MIMIC)-IV database. METHODS: We used the LASSO regression and the Boruta algorithm for feature selection. The dataset was split into training and test sets at a 7:3 ratio for constructing machine learning models. We employed logistic regression and nine other machine learning algorithms to build the prediction model. Restricted cubic spline (RCS) was used to analyse the linear relationship between the Braden score and the outcome, whereas the SHAP (Shapley additive explanations) method was applied to visualise the model's characteristics. RESULTS: This study compared the predictive ability of the Braden Scale with other scoring systems (SOFA, APSIII, Charlson, SAPSII). The results showed that the Braden Scale model had the highest performance, and SHAP analysis indicated that the Braden Scale is an important influencing factor for the risk of 90-day mortality in the ICU. The restricted cubic spline curve demonstrated a significant negative correlation between the Braden Scale and mortality. Subgroup analysis showed no significant interaction effects among subgroups except for age. CONCLUSIONS: The machine learning-enhanced Braden Scale has been developed to forecast the 90-day mortality risk for ICU patients suffering from pressure ulcers, and its efficacy as a clinically reliable tool has been substantiated. PATIENT OR PUBLIC CONTRIBUTION: Patients or public members were not directly involved in this study.

23A Rare Combination of Dermatomyositis, Interstitial Pneumonia, and Lung Cancer in a Patient Treated with Immunosuppressive Therapy and Chemotherapy.PubMed

Wakako Daido, Masahiro Yamasaki, Yuka Morio, et al.
Intern Med. 2018 Mar 15;57(6):849-853. doi: 10.2169/internalmedicine.9642-17. Epub 2017 Dec 8.
We herein report the rare case of co-occurring dermatomyositis (DM), interstitial pneumonia (IP), and lung cancer in a 59-year-old man. Computed tomography (CT) and positron emission tomography-CT showed the presence of a left lung tumor with IP, which was diagnosed as lung adenocarcinoma by a CT-guided tumor biopsy. We diagnosed DM based on the presence of myalgia, Gottron's papules, and anti-aminoacyl-tRNA synthetase antibody positivity in the patient. Co-occurrence of the above-mentioned three diseases is rare, and acute exacerbation of IP is a major cause of death in such cases. These patients can be treated with immunosuppressive therapy followed by chemotherapy.

24Summary of the Japanese Respiratory Society statement for the treatment of lung cancer with comorbid interstitial pneumonia.PubMed

Takashi Ogura, Nagio Takigawa, Keisuke Tomii, et al.
Respir Investig. 2019 Nov;57(6):512-533. doi: 10.1016/j.resinv.2019.06.001. Epub 2019 Aug 1.
Dramatic progress in targeted therapy and immunotherapy has been changing clinical practices in lung cancer. With the accumulation of clinical practice, it has become clear that pre-existing interstitial pneumonia (IP) could be a risk factor for drug-induced lung injury, which has enhanced awareness regarding the difficulty in treating lung cancer with comorbid IP. Unfortunately, there is only low-grade evidence in the field of lung cancer with comorbid IP, because almost all clinical trials exclude such patients. There have been very few specialized clinical trials for patients with lung cancer and underlying IPs thus far. Therefore, it is necessary to treat such cases empirically or to give up on the treatment itself. Considering these circumstances, establishing how to treat lung cancer with comorbid IP is an urgent issue. This paper is a summary of the official statement reported by the Diffuse Lung Disease/Thoracic Oncology Assembly and the Japanese Respiratory Society (JRS) in 2017, which attempts to approach lung cancer with comorbid IP systematically.

25Clinical spectrum of Chinese hospitalized lung cancer patients with concomitant interstitial lung disease: before and after the new era of LC treatment.PubMed

Ruxuan Chen, Chi Shao, Xiangning Liu, et al.
Clin Exp Med. 2023 Oct;23(6):2321-2330. doi: 10.1007/s10238-023-00999-1. Epub 2023 Jan 30.
This study aimed to explore the general characteristics and spectrum of hospitalized Chinese patients suffering from lung cancer with concomitant interstitial lung disease (LC-ILD). Furthermore, we compared their features before and after the period of immunotherapy for lung cancer. A retrospective analysis of the clinical characteristics of hospitalized LC patients with definite pathological diagnoses was performed from 2014 to 2021. ILD was defined after the review of chest CT imaging. There were 13,085 hospitalized LC patients. Among them, 509 patients (3.89%) had 551 cases of ILD. There were variable underlying causes of ILD, including idiopathic interstitial pneumonia (360 patients), LC treatment-associated ILD (134 cases), and connective tissue disease-associated ILD (55 patients). Although most LC-ILD patients were suffering from adenocarcinoma (204/40.1%), SCLC patients were prone to concomitant ILD (10.8% of all SCLC cases), followed by SCC (9.6% of all SCC cases). All but 10 LC-ILD patients received anti-LC treatment; however, only 39 (10.8%) LC-IIP patients received anti-ILD treatment. There were more LC-ILD patients in the 2018-2021 group than in the 2014-2017 group (5.16% vs. 2.03%, p < 0.001). The underlying causes of ILD were significantly different between the 2018-2021 group and the 2014-2017 group (p < 0.001). After adjusting for the number of hospitalized patients having the same LC pathological pattern, SCLC was determined to be the most likely to be concomitant with ILD, followed by SCC. Most LC-ILD patients were scheduled for anti-LC therapy; however, treatments for concomitant IIP were usually ignored. LC treatment-associated ILD should receive more attention than before.

26Independent association of metabolic tumor response on FDG-PET with pulmonary toxicity following risk-adaptive chemoradiation for unresectable non-small cell lung cancer: Inherent radiosensitivity or immune response?PubMed

Evan D H Gates, Daniel S Hippe, Hubert J Vesselle, et al.
Radiother Oncol. 2023 Aug;185:109720. doi: 10.1016/j.radonc.2023.109720. Epub 2023 May 25.
BACKGROUND: In the context of a phase II trial of risk-adaptive chemoradiation, we evaluated whether tumor metabolic response could serve as a correlate of treatment sensitivity and toxicity. METHODS: Forty-five patients with AJCCv7 stage IIB-IIIB NSCLC enrolled on the FLARE-RT phase II trial (NCT02773238). [18F]fluorodeoxyglucose (FDG) PET-CT images were acquired prior to treatment and after 24 Gy during week 3. Patients with unfavorable on-treatment tumor response received concomitant boosts to 74 Gy total over 30 fractions rather than standard 60 Gy. Metabolic tumor volume and mean standardized uptake value (SUVmean) were calculated semi-automatically. Risk factors of pulmonary toxicity included concurrent chemotherapy regimen, adjuvant anti-PDL1 immunotherapy, and lung dosimetry. Incidence of CTCAE v4 grade 2+ pneumonitis was analyzed using the Fine-Gray method with competing risks of metastasis or death. Peripheral germline DNA microarray sequencing measured predefined candidate genes from distinct pathways: 96 DNA repair, 53 immunology, 38 oncology, 27 lung biology. RESULTS: Twenty-four patients received proton therapy, 23 received ICI, 26 received carboplatin-paclitaxel, and 17 pneumonitis events were observed. Pneumonitis risk was significantly higher for patients with COPD (HR 3.78 [1.48, 9.60], p = 0.005), those treated with immunotherapy (HR 2.82 [1.03, 7.71], p = 0.043) but not with carboplatin-paclitaxel (HR 1.98 [0.71, 5.54], p = 0.19). Pneumonitis rates were similar among selected patients receiving 74 Gy radiation vs 60 Gy (p = 0.33), proton therapy vs photon (p = 0.60), or with higher lung dosimetric V20 (p = 0.30). Patients in the upper quartile decrease in SUVmean (>39.7%) were at greater risk for pneumonitis (HR 4.00 [1.54, 10.44], p = 0.005) and remained significant in multivariable analysis (HR 3.34 [1.23, 9.10], p = 0.018). Germline DNA gene alterations in immunology pathways were most frequently associated with pneumonitis. CONCLUSION: Tumor metabolic response as measured by mean SUV is associated with increased pneumonitis risk in a clinical trial cohort of NSCLC patients independent of treatment factors. This may be partially attributed to patient-specific differences in immunogenicity.

27Thyroxine alleviates interstitial lung disease induced by combined radiotherapy and immunotherapy.PubMed

You Mo, Yiwei Qin, Pengwei Li, et al.
Cancer Lett. 2025 Apr 10;615:217504. doi: 10.1016/j.canlet.2025.217504. Epub 2025 Jan 27.
Immune checkpoint blockade (ICB) combined with radiotherapy (RT) has improved patients survival, but also increased the risk of pulmonary adverse effects (AEs). Therefore, to explore potential drug targets for interstitial lung disease (ILD). We investigated the interaction of ICB and RT in pulmonary AEs using the disproportionality analysis and COX regression. Genome-wide association studies, transcriptome analysis, and vivo models highlighted the role of programmed death-ligand-1 (PD-L1) in ILD. Mendelian randomization analysis identified drug targets. Clinical data on pulmonary AEs in patients with non-small cell lung cancer (NSCLC) were used to validate the finding. Herein, we confirmed ICB combined with RT posed the highest risk of pulmonary AEs, with reporting odds ratio of interaction effect was 3.727. Anti-PD-L1 posed a greater risk of pulmonary AEs compared to anti-PD-1 (Hazard Ratio = 9.355) or anti-cytotoxic lymphocyte antigen-4 (CTLA-4) (Hazard Ratio = 6.985). In vivo study, PD-L1 expression in radiation induced lung injury negatively correlated with collagen deposition. Notably, thyroid hormone receptors were identified as drug targets for ILD. Our results demonstrate that thyroxine exerts a significant inhibitory effect on pulmonary fibrosis progression. Clinical evidence robustly supports the correlation between thyroid function and ILD. These findings highlight the importance of PD-L1 and thyroxine in understanding and managing ILD.

28Integration of dosimetric parameters, clinical factors, and radiomics to predict symptomatic radiation pneumonitis in lung cancer patients undergoing combined immunotherapy and radiotherapy.PubMed

Tingting Nie, Zien Chen, Jun Cai, et al.
Radiother Oncol. 2024 Jan;190:110047. doi: 10.1016/j.radonc.2023.110047. Epub 2023 Dec 7.
PURPOSE: This study aimed to combine clinical/dosimetric factors and handcrafted/deep learning radiomic features to establish a predictive model for symptomatic (grade ≥ 2) radiation pneumonitis (RP) in lung cancer patients who received immunotherapy followed by radiotherapy. MATERIALS AND METHODS: This study retrospectively collected data of 73 lung cancer patients with prior receipt of ICIs who underwent thoracic radiotherapy (TRT). Of these 73 patients, 41 (56.2 %) developed symptomatic grade ≥ 2 RP. RP was defined per multidisciplinary clinician consensus using CTCAE v5.0. Regions of interest (ROIs) (from radiotherapy planning CT images) utilized herein were gross tumor volume (GTV), planning tumor volume (PTV), and PTV-GTV. Clinical/dosimetric (mean lung dose and V5-V30) parameters were collected, and 107 handcrafted radiomic (HCR) features were extracted from each ROI. Deep learning-based radiomic (DLR) features were also extracted based on pre-trained 3D residual network models. HCR models, Fusion HCR model, Fusion HCR + ResNet models, and Fusion HCR + ResNet + Clinical models were built and compared using the receiver operating characteristic (ROC) curve with measurement of the area under the curve (AUC). Five-fold cross-validation was performed to avoid model overfitting. RESULTS: HCR models across various ROIs and the Fusion HCR model showed good predictive ability with AUCs from 0.740 to 0.808 and 0.740-0.802 in the training and testing cohorts, respectively. The addition of DLR features improved the effectiveness of HCR models (AUCs from 0.826 to 0.898 and 0.821-0.898 in both respective cohorts). The best performing prediction model (HCR + ResNet + Clinical) combined HCR & DLR features with 7 clinical/dosimetric characteristics and achieved an average AUC of 0.936 and 0.946 in both respective cohorts. CONCLUSIONS: In patients undergoing combined immunotherapy/RT for lung cancer, integrating clinical/dosimetric factors and handcrafted/deep learning radiomic features can offer a high predictive capacity for RP, and merits further prospective validation.

29Radiation pneumonitis prediction after stereotactic body radiation therapy based on 3D dose distribution: dosiomics and/or deep learning-based radiomics features.PubMed

Ying Huang, Aihui Feng, Yang Lin, et al.
Radiat Oncol. 2022 Nov 17;17(1):188. doi: 10.1186/s13014-022-02154-8.
BACKGROUND: This study was designed to establish radiation pneumonitis (RP) prediction models using dosiomics and/or deep learning-based radiomics (DLR) features based on 3D dose distribution. METHODS: A total of 140 patients with non-small cell lung cancer who received stereotactic body radiation therapy (SBRT) were retrospectively included in this study. These patients were randomly divided into the training (n = 112) and test (n = 28) sets. Besides, 107 dosiomics features were extracted by Pyradiomics, and 1316 DLR features were extracted by ResNet50. Feature visualization was performed based on Spearman's correlation coefficients, and feature selection was performed based on the least absolute shrinkage and selection operator. Three different models were constructed based on random forest, including (1) a dosiomics model (a model constructed based on dosiomics features), (2) a DLR model (a model constructed based on DLR features), and (3) a hybrid model (a model constructed based on dosiomics and DLR features). Subsequently, the performance of these three models was compared with receiver operating characteristic curves. Finally, these dosiomics and DLR features were analyzed with Spearman's correlation coefficients. RESULTS: In the training set, the area under the curve (AUC) of the dosiomics, DLR, and hybrid models was 0.9986, 0.9992, and 0.9993, respectively; the accuracy of these three models was 0.9643, 0.9464, and 0.9642, respectively. In the test set, the AUC of these three models was 0.8462, 0.8750, and 0.9000, respectively; the accuracy of these three models was 0.8214, 0.7857, and 0.8571, respectively. The hybrid model based on dosiomics and DLR features outperformed other two models. Correlation analysis between dosiomics features and DLR features showed weak correlations. The dosiomics features that correlated DLR features with the Spearman's rho |ρ| ≥ 0.8 were all first-order features. CONCLUSION: The hybrid features based on dosiomics and DLR features from 3D dose distribution could improve the performance of RP prediction after SBRT.

30Predicting radiation pneumonitis in locally advanced stage II-III non-small cell lung cancer using machine learning.PubMed

José Marcio Luna, Hann-Hsiang Chao, Eric S Diffenderfer, et al.
Radiother Oncol. 2019 Apr;133:106-112. doi: 10.1016/j.radonc.2019.01.003. Epub 2019 Jan 23.
BACKGROUND AND PURPOSE: Radiation pneumonitis (RP) is a radiotherapy dose-limiting toxicity for locally advanced non-small cell lung cancer (LA-NSCLC). Prior studies have proposed relevant dosimetric constraints to limit this toxicity. Using machine learning algorithms, we performed analyses of contributing factors in the development of RP to uncover previously unidentified criteria and elucidate the relative importance of individual factors. MATERIALS AND METHODS: We evaluated 32 clinical features per patient in a cohort of 203 stage II-III LA-NSCLC patients treated with definitive chemoradiation to a median dose of 66.6 Gy in 1.8 Gy daily fractions at our institution from 2008 to 2016. Of this cohort, 17.7% of patients developed grade ≥2 RP. Univariate analysis was performed using trained decision stumps to individually analyze statistically significant predictors of RP and perform feature selection. Applying Random Forest, we performed multivariate analysis to assess the combined performance of important predictors of RP. RESULTS: On univariate analysis, lung V20, lung mean, lung V10 and lung V5 were found to be significant RP predictors with the greatest balance of specificity and sensitivity. On multivariate analysis, Random Forest (AUC = 0.66, p = 0.0005) identified esophagus max (20.5%), lung V20 (16.4%), lung mean (15.7%) and pack-year (14.9%) as the most common primary differentiators of RP. CONCLUSIONS: We highlight Random Forest as an accurate machine learning method to identify known and new predictors of symptomatic RP. Furthermore, this analysis confirms the importance of lung V20, lung mean and pack-year as predictors of RP while also introducing esophagus max as an important RP predictor.

31Radiation pneumonitis prediction model with integrating multiple dose-function features on 4DCT ventilation images.PubMed

Yoshiyuki Katsuta, Noriyuki Kadoya, Tomohiro Kajikawa, et al.
Phys Med. 2023 Jan;105:102505. doi: 10.1016/j.ejmp.2022.11.009. Epub 2022 Dec 17.
PURPOSE: Radiation pneumonitis (RP) is dose-limiting toxicity for non-small-cell cancer (NSCLC). This study developed an RP prediction model by integrating dose-function features from computed four-dimensional computed tomography (4DCT) ventilation using the least absolute shrinkage and selection operator (LASSO). METHODS: Between 2013 and 2020, 126 NSCLC patients were included in this study who underwent a 4DCT scan to calculate ventilation images. We computed two sets of candidate dose-function features from (1) the percentage volume receiving > 20 Gy or the mean dose on the functioning zones determined with the lower cutoff percentile ventilation value, (2) the functioning zones determined with lower and upper cutoff percentile ventilation value using 4DCT ventilation images. An RP prediction model was developed by LASSO while simultaneously determining the regression coefficient and feature selection through fivefold cross-validation. RESULTS: We found 39.3 % of our patients had a ≥ grade 2 RP. The mean area under the curve (AUC) values for the developed models using clinical, dose-volume, and dose-function features with a lower cutoff were 0.791, and the mean AUC values with lower and upper cutoffs were 0.814. The relative regression coefficient (RRC) on dose-function features with upper and lower cutoffs revealed a relative impact of dose to each functioning zone to RP. RRCs were 0.52 for the mean dose on the functioning zone, with top 20 % of all functioning zone was two times greater than that of 0.19 for these with 60 %-80 % and 0.17 with 40 %-60 % (P < 0.01). CONCLUSIONS: The introduction of dose-function features computed from functioning zones with lower and upper cutoffs in a machine learning framework can improve RP prediction. The RRC given by LASSO using dose-function features allows for the quantification of the RP impact of dose on each functioning zones and having the potential to support treatment planning on functional image-guided radiotherapy.

32Prediction of long-term mortality by using machine learning models in Chinese patients with connective tissue disease-associated interstitial lung disease.PubMed

Di Sun, Yu Wang, Qing Liu, et al.
Respir Res. 2022 Jan 7;23(1):4. doi: 10.1186/s12931-022-01925-x.
BACKGROUND: The exact risk assessment is crucial for the management of connective tissue disease-associated interstitial lung disease (CTD-ILD) patients. In the present study, we develop a nomogram to predict 3‑ and 5-year mortality by using machine learning approach and test the ILD-GAP model in Chinese CTD-ILD patients. METHODS: CTD-ILD patients who were diagnosed and treated at the First Affiliated Hospital of Zhengzhou University were enrolled based on a prior well-designed criterion between February 2011 and July 2018. Cox regression with the least absolute shrinkage and selection operator (LASSO) was used to screen out the predictors and generate a nomogram. Internal validation was performed using bootstrap resampling. Then, the nomogram and ILD-GAP model were assessed via likelihood ratio testing, Harrell's C index, integrated discrimination improvement (IDI), the net reclassification improvement (NRI) and decision curve analysis. RESULTS: A total of 675 consecutive CTD-ILD patients were enrolled in this study, during the median follow-up period of 50 (interquartile range, 38-65) months, 158 patients died (mortality rate 23.4%). After feature selection, 9 variables were identified: age, rheumatoid arthritis, lung diffusing capacity for carbon monoxide, right ventricular diameter, right atrial area, honeycombing, immunosuppressive agents, aspartate transaminase and albumin. A predictive nomogram was generated by integrating these variables, which provided better mortality estimates than ILD-GAP model based on the likelihood ratio testing, Harrell's C index (0.767 and 0.652 respectively) and calibration plots. Application of the nomogram resulted in an improved IDI (3- and 5-year, 0.137 and 0.136 respectively) and NRI (3- and 5-year, 0.294 and 0.325 respectively) compared with ILD-GAP model. In addition, the nomogram was more clinically useful revealed by decision curve analysis. CONCLUSIONS: The results from our study prove that the ILD-GAP model may exhibit an inapplicable role in predicting mortality risk in Chinese CTD-ILD patients. The nomogram we developed performed well in predicting 3‑ and 5-year mortality risk of Chinese CTD-ILD patients, but further studies and external validation will be required to determine the clinical usefulness of the nomogram.

33Multicenter development of a deep learning radiomics and dosiomics nomogram to predict radiation pneumonia risk in non-small cell lung cancer.PubMed

Xun Wang, Aiping Zhang, Huipeng Yang, et al.
Sci Rep. 2025 May 16;15(1):17106. doi: 10.1038/s41598-025-02045-4.
Radiation pneumonia (RP) is the most common side effect of chest radiotherapy, and can affect patients' quality of life. This study aimed to establish a combined model of radiomics, dosiomics, deep learning (DL) based on simulated location CT and dosimetry images combining with clinical parameters to improve the predictive ability of ≥ 2 grade RP (RP2) in patients with non-small cell lung cancer (NSCLC). This study retrospectively collected 245 patients with NSCLC who received radiotherapy from three hospitals. 162 patients from Hospital I were randomly divided into training cohort and internal validation cohort according to 7:3. 83 patients from two other hospitals served as an external validation cohort. Multivariate analysis was used to screen independent clinical predictors and establish clinical model (CM). The radiomic and dosiomics (RD) features and DL features were extracted from simulated location CT and dosimetry images based on the region of interest (ROI) of total lung-PTV (TL-PTV). The features screened by the t-test and least absolute shrinkage and selection operator (LASSO) were used to construct the RD and DL model, and RD-score and DL-score were calculated. RD-score, DL-score and independent clinical features were combined to establish deep learning radiomics and dosiomics nomogram (DLRDN). The model performance was evaluated by area under the curve (AUC). Three clinical factors, including V20, V30, and mean lung dose (MLD), were used to establish the CM. 7 RD features including 4 radiomics features and 3 dosiomics features were selected to establish RD model. 10 DL features were selected to establish DL model. Among the different models, DLRDN showed the best predictions, with the AUCs of 0.891 (0.826-0.957), 0.825 (0.693-0.957), and 0.801 (0.698-0.904) in the training cohort, internal validation cohort and external validation cohort, respectively. DCA showed that DLRDN had a higher overall net benefit than other models. The calibration curve showed that the predicted value of DLRDN was in good agreement with the actual value. Overall, radiomics, dosiomics, and DL features based on simulated location CT and dosimetry images have the potential to help predict RP2. The combination of multi-dimensional data produced the optimal predictive model, which could provide guidance for clinicians.

34Machine learning-based techniques to improve lung transplantation outcomes and complications: a systematic review.PubMed

Marsa Gholamzadeh, Hamidreza Abtahi, Reza Safdari
BMC Med Res Methodol. 2022 Dec 23;22(1):331. doi: 10.1186/s12874-022-01823-2.
BACKGROUND: Machine learning has been used to develop predictive models to support clinicians in making better and more reliable decisions. The high volume of collected data in the lung transplant process makes it possible to extract hidden patterns by applying machine learning methods. Our study aims to investigate the application of machine learning methods in lung transplantation. METHOD: A systematic search was conducted in five electronic databases from January 2000 to June 2022. Then, the title, abstracts, and full text of extracted articles were screened based on the PRISMA checklist. Then, eligible articles were selected according to inclusion criteria. The information regarding developed models was extracted from reviewed articles using a data extraction sheet. RESULTS: Searches yielded 414 citations. Of them, 136 studies were excluded after the title and abstract screening. Finally, 16 articles were determined as eligible studies that met our inclusion criteria. The objectives of eligible articles are classified into eight main categories. The applied machine learning methods include the Support vector machine (SVM) (n = 5, 31.25%) technique, logistic regression (n = 4, 25%), Random Forests (RF) (n = 4, 25%), Bayesian network (BN) (n = 3, 18.75%), linear regression (LR) (n = 3, 18.75%), Decision Tree (DT) (n = 3, 18.75%), neural networks (n = 3, 18.75%), Markov Model (n = 1, 6.25%), KNN (n = 1, 6.25%), K-means (n = 1, 6.25%), Gradient Boosting trees (XGBoost) (n = 1, 6.25%), and Convolutional Neural Network (CNN) (n = 1, 6.25%). Most studies (n = 11) employed more than one machine learning technique or combination of different techniques to make their models. The data obtained from pulmonary function tests were the most used as input variables in predictive model development. Most studies (n = 10) used only post-transplant patient information to develop their models. Also, UNOS was recognized as the most desirable data source in the reviewed articles. In most cases, clinicians succeeded to predict acute diseases incidence after lung transplantation (n = 4) or estimate survival rate (n = 4) by developing machine learning models. CONCLUSION: The outcomes of these developed prediction models could aid clinicians to make better and more reliable decisions by extracting new knowledge from the huge volume of lung transplantation data.

35Development of a novel combined nomogram model integrating deep learning-pathomics, radiomics and immunoscore to predict postoperative outcome of colorectal cancer lung metastasis patients.PubMed

Renjie Wang, Weixing Dai, Jing Gong, et al.
J Hematol Oncol. 2022 Jan 24;15(1):11. doi: 10.1186/s13045-022-01225-3.
Limited previous studies focused on the death and progression risk stratification of colorectal cancer (CRC) lung metastasis patients. The aim of this study is to construct a nomogram model combing machine learning-pathomics, radiomics features, Immunoscore and clinical factors to predict the postoperative outcome of CRC patients with lung metastasis. In this study, a total of 103 CRC patients having metastases limited to lung and undergoing radical lung resection were identified. Patch-level convolutional neural network training in weakly supervised manner was used to perform whole slides histopathological images survival analysis. Synthetic minority oversampling technique and support vector machine classifier were used to identify radiomics features and build predictive signature. The Immunoscore for each patient was calculated from the density of CD3+ and CD8+ cells at the invasive margin and the center of metastatic tumor which were assessed on consecutive sections of automated digital pathology. Finally, pathomics and radiomics signatures were successfully developed to predict the overall survival (OS) and disease free survival (DFS) of patients. The predicted pathomics and radiomics scores are negatively correlated with Immunoscore and they are three independent prognostic factors for OS and DFS prediction. The combined nomogram showed outstanding performance in predicting OS (AUC = 0.860) and DFS (AUC = 0.875). The calibration curve and decision curve analysis demonstrated the considerable clinical usefulness of the combined nomogram. Taken together, the developed nomogram model consisting of machine learning-pathomics signature, radiomics signature, Immunoscore and clinical features could be reliable in predicting postoperative OS and DFS of colorectal lung metastasis patients.

36Shallow and deep learning classifiers in medical image analysis.PubMed

Francesco Prinzi, Tiziana Currieri, Salvatore Gaglio, et al.
Eur Radiol Exp. 2024 Mar 5;8(1):26. doi: 10.1186/s41747-024-00428-2.
An increasingly strong connection between artificial intelligence and medicine has enabled the development of predictive models capable of supporting physicians' decision-making. Artificial intelligence encompasses much more than machine learning, which nevertheless is its most cited and used sub-branch in the last decade. Since most clinical problems can be modeled through machine learning classifiers, it is essential to discuss their main elements. This review aims to give primary educational insights on the most accessible and widely employed classifiers in radiology field, distinguishing between "shallow" learning (i.e., traditional machine learning) algorithms, including support vector machines, random forest and XGBoost, and "deep" learning architectures including convolutional neural networks and vision transformers. In addition, the paper outlines the key steps for classifiers training and highlights the differences between the most common algorithms and architectures. Although the choice of an algorithm depends on the task and dataset dealing with, general guidelines for classifier selection are proposed in relation to task analysis, dataset size, explainability requirements, and available computing resources. Considering the enormous interest in these innovative models and architectures, the problem of machine learning algorithms interpretability is finally discussed, providing a future perspective on trustworthy artificial intelligence.Relevance statement The growing synergy between artificial intelligence and medicine fosters predictive models aiding physicians. Machine learning classifiers, from shallow learning to deep learning, are offering crucial insights for the development of clinical decision support systems in healthcare. Explainability is a key feature of models that leads systems toward integration into clinical practice. Key points • Training a shallow classifier requires extracting disease-related features from region of interests (e.g., radiomics).• Deep classifiers implement automatic feature extraction and classification.• The classifier selection is based on data and computational resources availability, task, and explanation needs.

37Classification of malignant lung cancer using deep learning.PubMed

Vinod Kumar, Brijesh Bakariya
J Med Eng Technol. 2021 Feb;45(2):85-93. doi: 10.1080/03091902.2020.1853837. Epub 2021 Jan 15.
In the automatic detection of suspicious shaded regions on CT images derived from the LIDC-IDRI dataset, the diagnostic system plays a significant role. This paper introduces an automatic recognition method for lung nodules of the regions of concern (ROI). The lung regions are segmented from DICOM image size 512 × 512 by adding a median filter, Gaussian filter, Gabor filter and watershed algorithm. AlexNet uses 227 × 227 × 3 with "fc7" (fully connected) layers and GoogLeNet uses 224 × 224 × 3 with "pool5-drop 7 × 7 s1" layers. Here, the authors explain what is better about AlexNet and GoogLeNet through its performance analysis, feature extraction, classification, sensitivity, specificity, detection and false alarm rate with time complexity. A multi-class SVM classifier with 100% precision and specificity provided the best performance in deep learning neural networks.

38A systematic review on feature extraction methods and deep learning models for detection of cancerous lung nodules at an early stage -the recent trends and challenges.PubMed

Mathumetha Palani, Sivakumar Rajagopal, Anantha Krishna Chintanpalli
Biomed Phys Eng Express. 2024 Nov 20;11(1). doi: 10.1088/2057-1976/ad9154.
Lung cancer is one of the most common life-threatening worldwide cancers affecting both the male and the female populations. The appearance of nodules in the scan image is an early indication of the development of cancer cells in the lung. The Low Dose Computed Tomography screening technique is used for the early detection of cancer nodules. Therefore, with more Computed Tomography (CT) lung profiles, an automated lung nodule analysis system can be utilized through image processing techniques and neural network algorithms. A CT image of the lung consists of many elements such as blood vessels, ribs, nodules, sternum, bronchi and nodules. These nodules can be both benign and malignant, where the latter leads to lung cancer. Detecting them at an earlier stage can increase life expectancy by up to 5 to 10 years. To analyse only the nodules from the profile, the respected features are extracted using image processing techniques. Based on the review, textural features were the promising ones in medical image analysis and for solving computer vision problems. The importance of uncovering the hidden features allows Deep Learning algorithms (DL) to function better, especially in medical imaging, where accuracy has improved. The earlier detection of cancerous lung nodules is possible through the combination of multi-featured extraction and classification techniques using image data. This technique can be a breakthrough in the deep learning area by providing the appropriate features. One of the greatest challenges is the incorrect identification of malignant nodules results in a higher false positive rate during the prediction. The suitable features make the system more precise in prognosis. In this paper, the overview of lung cancer along with the publicly available datasets is discussed for the research purposes. They are mainly focused on the recent research that combines feature extraction and deep learning algorithms used to reduce the false positive rate in the automated detection of lung nodules. The primary objective of the paper is to provide the importance of textural features when combined with different deep-learning models. It gives insights into their advantages, disadvantages and limitations regarding possible research gaps. These papers compare the recent studies of deep learning models with and without feature extraction and conclude that DL models that include feature extraction are better than the others.

39Automated Lung Cancer Segmentation Using a PET and CT Dual-Modality Deep Learning Neural Network.PubMed

Siqiu Wang, Rebecca Mahon, Elisabeth Weiss, et al.
Int J Radiat Oncol Biol Phys. 2023 Feb 1;115(2):529-539. doi: 10.1016/j.ijrobp.2022.07.2312. Epub 2022 Aug 4.
PURPOSE: To develop an automated lung tumor segmentation method for radiation therapy planning based on deep learning and dual-modality positron emission tomography (PET) and computed tomography (CT) images. METHODS AND MATERIALS: A 3-dimensional (3D) convolutional neural network using inputs from diagnostic PETs and simulation CTs was constructed with 2 parallel convolution paths for independent feature extraction at multiple resolution levels and a single deconvolution path. At each resolution level, the extracted features from the convolution arms were concatenated and fed through the skip connections into the deconvolution path that produced the tumor segmentation. Our network was trained/validated/tested by a 3:1:1 split on 290 pairs of PET and CT images from patients with lung cancer treated at our clinic, with manual physician contours as the ground truth. A stratified training strategy based on the magnitude of the gross tumor volume (GTV) was investigated to improve performance, especially for small tumors. Multiple radiation oncologists assessed the clinical acceptability of the network-produced segmentations. RESULTS: The mean Dice similarity coefficient, Hausdorff distance, and bidirectional local distance comparing manual versus automated contours were 0.79 ± 0.10, 5.8 ± 3.2 mm, and 2.8 ± 1.5 mm for the unstratified 3D dual-modality model. Stratification delivered the best results when the model for the large GTVs (>25 mL) was trained with all-size GTVs and the model for the small GTVs (<25 mL) was trained with small GTVs only. The best combined Dice similarity coefficient, Hausdorff distance, and bidirectional local distance from the 2 stratified models on their corresponding test data sets were 0.83 ± 0.07, 5.9 ± 2.5 mm, and 2.8 ± 1.4 mm, respectively. In the multiobserver review, 91.25% manual versus 88.75% automatic contours were accepted or accepted with modifications. CONCLUSIONS: By using an expansive clinical PET and CT image database and a dual-modality architecture, the proposed 3D network with a novel GTVbased stratification strategy generated clinically useful lung cancer contours that were highly acceptable on physician review.

40Computer-aided diagnosis of ground glass pulmonary nodule by fusing deep learning and radiomics features.PubMed

Xianfang Hu, Jing Gong, Wei Zhou, et al.
Phys Med Biol. 2021 Mar 4;66(6):065015. doi: 10.1088/1361-6560/abe735.
OBJECTIVES: This study aims to develop a computer-aided diagnosis (CADx) scheme to classify between benign and malignant ground glass nodules (GGNs), and fuse deep leaning and radiomics imaging features to improve the classification performance. METHODS: We first retrospectively collected 513 surgery histopathology confirmed GGNs from two centers. Among these GGNs, 100 were benign and 413 were malignant. All malignant tumors were stage I lung adenocarcinoma. To segment GGNs, we applied a deep convolutional neural network and residual architecture to train and build a 3D U-Net. Then, based on the pre-trained U-Net, we used a transfer learning approach to build a deep neural network (DNN) to classify between benign and malignant GGNs. With the GGN segmentation results generated by 3D U-Net, we also developed a CT radiomics model by adopting a series of image processing techniques, i.e. radiomics feature extraction, feature selection, synthetic minority over-sampling technique, and support vector machine classifier training/testing, etc. Finally, we applied an information fusion method to fuse the prediction scores generated by DNN based CADx model and CT-radiomics based model. To evaluate the proposed model performance, we conducted a comparison experiment by testing on an independent testing dataset. RESULTS: Comparing with DNN model and radiomics model, our fusion model yielded a significant higher area under a receiver operating characteristic curve (AUC) value of 0.73 ± 0.06 (P < 0.01). The fusion model generated an accuracy of 75.6%, F1 score of 84.6%, weighted average F1 score of 70.3%, and Matthews correlation coefficient of 43.6%, which were higher than the DNN model and radiomics model individually. CONCLUSIONS: Our experimental results demonstrated that (1) applying a CADx scheme was feasible to diagnosis of early-stage lung adenocarcinoma, (2) deep image features and radiomics features provided complementary information in classifying benign and malignant GGNs, and (3) it was an effective way to build DNN model with limited dataset by using transfer learning. Thus, to build a robust image analysis based CADx model, one can combine different types of image features to decode the imaging phenotypes of GGN.

41Deep Learning-based Segmentation of Computed Tomography Scans Predicts Disease Progression and Mortality in Idiopathic Pulmonary Fibrosis.PubMed

Muhunthan Thillai, Justin M Oldham, Alessandro Ruggiero, et al.
Am J Respir Crit Care Med. 2024 Aug 15;210(4):465-472. doi: 10.1164/rccm.202311-2185OC.
Despite evidence demonstrating a prognostic role for computed tomography (CT) scans in idiopathic pulmonary fibrosis (IPF), image-based biomarkers are not routinely used in clinical practice or trials. To develop automated imaging biomarkers using deep learning-based segmentation of CT scans. We developed segmentation processes for four anatomical biomarkers, which were applied to a unique cohort of treatment-naive patients with IPF enrolled in the PROFILE (Prospective Observation of Fibrosis in the Lung Clinical Endpoints) study and tested against a further United Kingdom cohort. The relationships among CT biomarkers, lung function, disease progression, and mortality were assessed. Data from 446 PROFILE patients were analyzed. Median follow-up duration was 39.1 months (interquartile range, 18.1-66.4 mo), with a cumulative incidence of death of 277 (62.1%) over 5 years. Segmentation was successful on 97.8% of all scans, across multiple imaging vendors, at slice thicknesses of 0.5-5 mm. Of four segmentations, lung volume showed the strongest correlation with FVC ( = 0.82; < 0.001). Lung, vascular, and fibrosis volumes were consistently associated across cohorts with differential 5-year survival, which persisted after adjustment for baseline gender, age, and physiology score. Lower lung volume (hazard ratio [HR], 0.98 [95% confidence interval (CI), 0.96-0.99]; = 0.001), increased vascular volume (HR, 1.30 [95% CI, 1.12-1.51]; = 0.001), and increased fibrosis volume (HR, 1.17 [95% CI, 1.12-1.22]; < 0.001) were associated with reduced 2-year progression-free survival in the pooled PROFILE cohort. Longitudinally, decreasing lung volume (HR, 3.41 [95% CI, 1.36-8.54]; = 0.009) and increasing fibrosis volume (HR, 2.23 [95% CI, 1.22-4.08]; = 0.009) were associated with differential survival. Automated models can rapidly segment IPF CT scans, providing prognostic near and long-term information, which could be used in routine clinical practice or as key trial endpoints.

42Lung cancer detection and classification using optimized CNN features and Squeeze-Inception-ResNeXt model.PubMed

Geethu Lakshmi G, P Nagaraj
Comput Biol Chem. 2025 Aug;117:108437. doi: 10.1016/j.compbiolchem.2025.108437. Epub 2025 Mar 25.
Lung cancer, with its high mortality rate, is one of the deadliest diseases globally. The alarming increase in lung cancer deaths and its widespread prevalence have led to the development of various cancer control research and early detection methods aimed at reducing mortality rates. Effective diagnostic techniques are crucial for lowering lung cancer incidence, as early detection significantly impacts treatment success. Human error can often impede accurate identification of lung nodules, in which Computer-Aided Diagnostic (CAD) systems are utilized. These systems help radiologists by automating diagnostic processes and improving accuracy of detecting and classifying malignancies. This paper aims to develop a deep learning approach for classifying lung diseases using chest Computed Tomography (CT) scan images. The approach starts with image pre-processing, including color space conversion, data augmentation, resizing, and normalization. Feature extraction is carried out using a Convolutional Neural Network (CNN) optimized with Slime Mould Algorithm (SMA). For classification, a novel approach combining Squeeze-Inception V3 with ResNeXt, referred to as Squeeze-Inception-ResNeXt, is proposed. The Squeeze-Inception-ResNeXt model benefits from reduced computational cost while maintaining high performance in classifying lung diseases. This model categorizes lung diseases into Adenocarcinoma, Large Cell Carcinoma, and Squamous Cell Carcinoma. Additionally, SMA is utilized in training the Squeeze-Inception-ResNeXt model. Experimental results show that Squeeze-Inception-ResNeXt surpasses traditional models, with an accuracy of 97.7 %, sensitivity of 98.1 %, and specificity of 97.4 %.

43Glaucoma detection: Binocular approach and clinical data in machine learning.PubMed

Oleksandr Kovalyk-Borodyak, Juan Morales-Sánchez, Rafael Verdú-Monedero, et al.
Artif Intell Med. 2025 Feb;160:103050. doi: 10.1016/j.artmed.2024.103050. Epub 2024 Dec 12.
In this work, we present a multi-modal machine learning method to automate early glaucoma diagnosis. The proposed methodology introduces two novel aspects for automated diagnosis not previously explored in the literature: simultaneous use of ocular fundus images from both eyes and integration with the patient's additional clinical data. We begin by establishing a baseline, termed monocular mode, which adheres to the traditional approach of considering the data from each eye as a separate instance. We then explore the binocular mode, investigating how combining information from both eyes of the same patient can enhance glaucoma diagnosis accuracy. This exploration employs the PAPILA dataset, comprising information from both eyes, clinical data, ocular fundus images, and expert segmentation of these images. Additionally, we compare two image-derived data modalities: direct ocular fundus images and morphological data from manual expert segmentation. Our method integrates Gradient-Boosted Decision Trees (GBDT) and Convolutional Neural Networks (CNN), specifically focusing on the MobileNet, VGG16, ResNet-50, and Inception models. SHAP values are used to interpret GBDT models, while the Deep Explainer method is applied in conjunction with SHAP to analyze the outputs of convolutional-based models. Our findings show the viability of considering both eyes, which improves the model performance. The binocular approach, incorporating information from morphological and clinical data yielded an AUC of 0.796 (±0.003 at a 95% confidence interval), while the CNN, using the same approach (both eyes), achieved an AUC of 0.764 (±0.005 at a 95% confidence interval).

44Multi-view contrastive learning and symptom extraction insights for medical report generation.PubMed

Qi Bai, Xiaodi Zou, Ahmad Alhaskawi, et al.
Sci Rep. 2025 May 23;15(1):17991. doi: 10.1038/s41598-025-00570-w.
The task of generating medical reports automatically is of paramount importance in modern healthcare, offering a substantial reduction in the workload of radiologists and accelerating the processes of clinical diagnosis and treatment. Current challenges include handling limited sample sizes and interpreting intricate multi-modal and multi-view medical data. In order to improve the accuracy and efficiency for radiologists, we conducted this investigation. This study aims to present a novel methodology for medical report generation that leverages Multi-View Contrastive Learning (MVCL) applied to MRI data, combined with a Symptom Consultant (SC) for extracting medical insights, to improve the quality and efficiency of automated medical report generation. We introduce an advanced MVCL framework that maximizes the potential of multi-view MRI data to enhance visual feature extraction. Alongside, the SC component is employed to distill critical medical insights from symptom descriptions. These components are integrated within a transformer decoder architecture, which is then applied to the Deep Wrist dataset for model training and evaluation. Our experimental analysis on the Deep Wrist dataset reveals that our proposed integration of MVCL and SC significantly outperforms the baseline model in terms of accuracy and relevance of the generated medical reports. The results indicate that our approach is particularly effective in capturing and utilizing the complex information inherent in multi-modal and multi-view medical datasets. The combination of MVCL and SC constitutes a powerful approach to medical report generation, addressing the existing challenges in the field. The demonstrated superiority of our model over traditional methods holds promise for substantial improvements in clinical diagnosis and automated report generation, indicating a significant stride forward in medical technology.

45Multimodality Fusion Method Based on Multiview Subspace Clustering for Pulmonary Embolism Diagnosis.PubMed

Peijun Hu, Qianqian Qi, Yanxia Zhao, et al.
Stud Health Technol Inform. 2024 Jan 25;310:916-920. doi: 10.3233/SHTI231098.
Pulmonary embolism (PE) is an important clinical disorder that will result in lung tissue damage or low blood oxygen levels, which need early diagnosis and timely treatment. While computed tomographic pulmonary angiography (CTPA) is the gold standard to diagnose PE, previous studies have verified the effectiveness of combing CTPA and EMR data in computer-aided PE detection or diagnosis. In this paper, we proposed a multimodality fusion method based on multi-view subspace clustering guided feature selection (MSCUFS). The extracted high-dimensional image and EMR features are firstly selected and fused by the MSCUFS, and then are feed into different machine learning models with different fusion strategy to construct the PE classifier. The experiment results showed that the joint fusion strategy with MSCUFS achieved best AUROC of 0.947, surpassing other early fusion and late fusion models. The comparison between single modality and multimodality also illustrated the effectiveness of the proposed method.

46Identification of Parkinson's disease using MRI and genetic data from the PPMI cohort: an improved machine learning fusion approach.PubMed

Yifeng Yang, Liangyun Hu, Yang Chen, et al.
Front Aging Neurosci. 2025 Feb 4;17:1510192. doi: 10.3389/fnagi.2025.1510192. eCollection 2025.
OBJECTIVE: This study aim to leverage advanced machine learning techniques to develop and validate novel MRI imaging features and single nucleotide polymorphism (SNP) gene data fusion methodologies to enhance the early identification and diagnosis of Parkinson's disease (PD). METHODS: We leveraged a comprehensive dataset from the Parkinson's Progression Markers Initiative (PPMI), which includes high-resolution neuroimaging data, genetic single-nucleotide polymorphism (SNP) profiles, and detailed clinical information from individuals with early-stage PD and healthy controls. Two multi-modal fusion strategies were used: feature-level fusion, where we employed a hybrid feature selection algorithm combining Fisher discriminant analysis, an ensemble Lasso (EnLasso) method, and partial least squares (PLS) regression to identify and integrate the most informative features from neuroimaging and genetic data; and decision-level fusion, where we developed an adaptive ensemble stacking (AE_Stacking) model to synergistically integrate the predictions from multiple base classifiers trained on individual modalities. RESULTS: The AE_Stacking model achieving the highest average balanced accuracy of 95.36% and an area under the receiver operating characteristic curve (AUC) of 0.974, significantly outperforming feature-level fusion and single-modal models ( < 0.05). Furthermore, by analyzing the features selected across multiple iterations of our models, we identified stable brain region features [lh 6r (FD) and rh 46 (GI)] and key genetic markers (rs356181 and rs2736990 SNPs within the SNCA gene region; rs213202 SNP within the VPS52 gene region), highlighting their potential as reliable early diagnostic indicators for the disease. CONCLUSION: The AE_Stacking model, trained on MRI and genetic data, demonstrates potential in distinguishing individuals with PD. Our findings enhance understanding of the disease and advance us toward the goal of precision medicine for neurodegenerative disorder.

47A review of medical image data augmentation techniques for deep learning applications.PubMed

Phillip Chlap, Hang Min, Nym Vandenberg, et al.
J Med Imaging Radiat Oncol. 2021 Aug;65(5):545-563. doi: 10.1111/1754-9485.13261. Epub 2021 Jun 19.
Research in artificial intelligence for radiology and radiotherapy has recently become increasingly reliant on the use of deep learning-based algorithms. While the performance of the models which these algorithms produce can significantly outperform more traditional machine learning methods, they do rely on larger datasets being available for training. To address this issue, data augmentation has become a popular method for increasing the size of a training dataset, particularly in fields where large datasets aren't typically available, which is often the case when working with medical images. Data augmentation aims to generate additional data which is used to train the model and has been shown to improve performance when validated on a separate unseen dataset. This approach has become commonplace so to help understand the types of data augmentation techniques used in state-of-the-art deep learning models, we conducted a systematic review of the literature where data augmentation was utilised on medical images (limited to CT and MRI) to train a deep learning model. Articles were categorised into basic, deformable, deep learning or other data augmentation techniques. As artificial intelligence models trained using augmented data make their way into the clinic, this review aims to give an insight to these techniques and confidence in the validity of the models produced.

48Data augmentation for medical imaging: A systematic literature review.PubMed

Fabio Garcea, Alessio Serra, Fabrizio Lamberti, et al.
Comput Biol Med. 2023 Jan;152:106391. doi: 10.1016/j.compbiomed.2022.106391. Epub 2022 Dec 9.
Recent advances in Deep Learning have largely benefited from larger and more diverse training sets. However, collecting large datasets for medical imaging is still a challenge due to privacy concerns and labeling costs. Data augmentation makes it possible to greatly expand the amount and variety of data available for training without actually collecting new samples. Data augmentation techniques range from simple yet surprisingly effective transformations such as cropping, padding, and flipping, to complex generative models. Depending on the nature of the input and the visual task, different data augmentation strategies are likely to perform differently. For this reason, it is conceivable that medical imaging requires specific augmentation strategies that generate plausible data samples and enable effective regularization of deep neural networks. Data augmentation can also be used to augment specific classes that are underrepresented in the training set, e.g., to generate artificial lesions. The goal of this systematic literature review is to investigate which data augmentation strategies are used in the medical domain and how they affect the performance of clinical tasks such as classification, segmentation, and lesion detection. To this end, a comprehensive analysis of more than 300 articles published in recent years (2018-2022) was conducted. The results highlight the effectiveness of data augmentation across organs, modalities, tasks, and dataset sizes, and suggest potential avenues for future research.

49Deep learning for categorization of endodontic lesion based on radiographic periapical index scoring system.PubMed

Navas P Moidu, Sidhartha Sharma, Amrita Chawla, et al.
Clin Oral Investig. 2022 Jan;26(1):651-658. doi: 10.1007/s00784-021-04043-y. Epub 2021 Jul 2.
OBJECTIVE: The study aimed to apply convolutional neural network (CNN) to score periapical lesion on an intraoral periapical radiograph (IOPAR) based on the periapical index (PAI) scoring system. MATERIALS AND METHODS: A total of 3000 periapical root areas (PRA) on 1950 digital IOPAR were pre-scored by three endodontists. This data was used to train the CNN model-"YOLO version 3." A total of 450 PRA was used for validation of the model. Data augmentation techniques and model optimization were applied. A total of 540 PRA on 250 digital IOPAR was used to test the performance of the CNN model. RESULTS: A total of 303 PRA (56.11%) exhibited true prediction. PAI score 1 showed the highest true prediction (90.9%). PAI scores 2 and 5 exhibited the least true prediction (30% each). PAI scores 3 and 4 had a true prediction of 60% and 71%, respectively. When the scores were dichotomized as healthy (PAI scores 1 and 2) and diseased (PAI score 3, 4, and 5), the model achieved a true prediction of 76.6% and 92%, respectively. The model exhibited a 92.1% sensitivity/recall, 76% specificity, 86.4% positive predictive value/precision, and 86.1% negative predictive value. The accuracy, F1 score, and Matthews correlation coefficient were 86.3%, 0.89, and 0.71, respectively. CONCLUSION: The CNN model trained on a limited amount of IOPAR data showed potential for PAI scoring of the periapical lesion on digital IOPAR. CLINICAL RELEVANCE: An automated system for PAI scoring is developed that would potentially benefit clinician and researchers.

50Domain-guided data augmentation for deep learning on medical imaging.PubMed

Chinmayee Athalye, Rima Arnaout
PLoS One. 2023 Mar 23;18(3):e0282532. doi: 10.1371/journal.pone.0282532. eCollection 2023.
While domain-specific data augmentation can be useful in training neural networks for medical imaging tasks, such techniques have not been widely used to date. Our objective was to test whether domain-specific data augmentation is useful for medical imaging using a well-benchmarked task: view classification on fetal ultrasound FETAL-125 and OB-125 datasets. We found that using a context-preserving cut-paste strategy, we could create valid training data as measured by performance of the resulting trained model on the benchmark test dataset. When used in an online fashion, models trained on this hybrid data performed similarly to those trained using traditional data augmentation (FETAL-125 F-score 85.33 ± 0.24 vs 86.89 ± 0.60, p-value 0.014; OB-125 F-score 74.60 ± 0.11 vs 72.43 ± 0.62, p-value 0.004). Furthermore, the ability to perform augmentations during training time, as well as the ability to apply chosen augmentations equally across data classes, are important considerations in designing a bespoke data augmentation. Finally, we provide open-source code to facilitate running bespoke data augmentations in an online fashion. Taken together, this work expands the ability to design and apply domain-guided data augmentations for medical imaging tasks.

51AI applications to medical images: From machine learning to deep learning.PubMed

Isabella Castiglioni, Leonardo Rundo, Marina Codari, et al.
Phys Med. 2021 Mar;83:9-24. doi: 10.1016/j.ejmp.2021.02.006. Epub 2021 Mar 1.
PURPOSE: Artificial intelligence (AI) models are playing an increasing role in biomedical research and healthcare services. This review focuses on challenges points to be clarified about how to develop AI applications as clinical decision support systems in the real-world context. METHODS: A narrative review has been performed including a critical assessment of articles published between 1989 and 2021 that guided challenging sections. RESULTS: We first illustrate the architectural characteristics of machine learning (ML)/radiomics and deep learning (DL) approaches. For ML/radiomics, the phases of feature selection and of training, validation, and testing are described. DL models are presented as multi-layered artificial/convolutional neural networks, allowing us to directly process images. The data curation section includes technical steps such as image labelling, image annotation (with segmentation as a crucial step in radiomics), data harmonization (enabling compensation for differences in imaging protocols that typically generate noise in non-AI imaging studies) and federated learning. Thereafter, we dedicate specific sections to: sample size calculation, considering multiple testing in AI approaches; procedures for data augmentation to work with limited and unbalanced datasets; and the interpretability of AI models (the so-called black box issue). Pros and cons for choosing ML versus DL to implement AI applications to medical imaging are finally presented in a synoptic way. CONCLUSIONS: Biomedicine and healthcare systems are one of the most important fields for AI applications and medical imaging is probably the most suitable and promising domain. Clarification of specific challenging points facilitates the development of such systems and their translation to clinical practice.

52Machine Learning Model Drift: Predicting Diagnostic Imaging Follow-Up as a Case Example.PubMed

Ronilda Lacson, Mahsa Eskian, Andro Licaros, et al.
J Am Coll Radiol. 2022 Oct;19(10):1162-1169. doi: 10.1016/j.jacr.2022.05.030. Epub 2022 Aug 16.
OBJECTIVE: Address model drift in a machine learning (ML) model for predicting diagnostic imaging follow-up using data augmentation with more recent data versus retraining new predictive models. METHODS: This institutional review board-approved retrospective study was conducted January 1, 2016, to December 31, 2020, at a large academic institution. A previously trained ML model was trained on 1,000 radiology reports from 2016 (old data). An additional 1,385 randomly selected reports from 2019 to 2020 (new data) were annotated for follow-up recommendations and randomly divided into two sets: training (n = 900) and testing (n = 485). Support vector machine and random forest (RF) algorithms were constructed and trained using 900 new data reports plus old data (augmented data, new models) and using only new data (new data, new models). The 2016 baseline model was used as comparator as is and trained with augmented data. Recall was compared with baseline using McNemar's test. RESULTS: Follow-up recommendations were contained in 11.3% of reports (157 or 1,385). The baseline model retrained with new data had precision = 0.83 and recall = 0.54; none significantly different from baseline. A new RF model trained with augmented data had significantly better recall versus the baseline model (0.80 versus 0.66, P = .04) and comparable precision (0.90 versus 0.86). DISCUSSION: ML methods for monitoring follow-up recommendations in radiology reports suffer model drift over time. A newly developed RF model achieved better recall with comparable precision versus simply retraining a previously trained original model with augmented data. Thus, regularly assessing and updating these models is necessary using more recent historical data.

53Deep multimodal learning for lymph node metastasis prediction of primary thyroid cancer.PubMed

Xinglong Wu, Mengying Li, Xin-Wu Cui, et al.
Phys Med Biol. 2022 Feb 1;67(3). doi: 10.1088/1361-6560/ac4c47.
. The incidence of primary thyroid cancer has risen steadily over the past decades because of overdiagnosis and overtreatment through the improvement in imaging techniques for screening, especially in ultrasound examination. Metastatic status of lymph nodes is important for staging the type of primary thyroid cancer. Deep learning algorithms based on ultrasound images were thus developed to assist radiologists on the diagnosis of lymph node metastasis. The objective of this study is to integrate more clinical context (e.g., health records and various image modalities) into, and explore more interpretable patterns discovered by, deep learning algorithms for the prediction of lymph node metastasis in primary thyroid cancer patients.. A deep multimodal learning network was developed in this study with a novel index proposed to compare the contribution of different modalities when making the predictions.. The proposed multimodal network achieved an average F1 score of 0.888 and an average area under the receiver operating characteristic curve (AUC) value of 0.973 in two independent validation sets, and the performance was significantly better than that of three single-modality deep learning networks. Moreover, among three modalities used in this study, the deep multimodal learning network relied generally more on image modalities than the data modality of clinic records when making the predictions.. Our work is beneficial to prospective clinic trials of radiologists on the diagnosis of lymph node metastasis in primary thyroid cancer, and will better help them understand how the predictions are made in deep multimodal learning algorithms.

54Using a Machine Learning Approach to Predict Outcomes after Radiosurgery for Cerebral Arteriovenous Malformations.PubMed

Eric Karl Oermann, Alex Rubinsteyn, Dale Ding, et al.
Sci Rep. 2016 Feb 9;6:21161. doi: 10.1038/srep21161.
Predictions of patient outcomes after a given therapy are fundamental to medical practice. We employ a machine learning approach towards predicting the outcomes after stereotactic radiosurgery for cerebral arteriovenous malformations (AVMs). Using three prospective databases, a machine learning approach of feature engineering and model optimization was implemented to create the most accurate predictor of AVM outcomes. Existing prognostic systems were scored for purposes of comparison. The final predictor was secondarily validated on an independent site's dataset not utilized for initial construction. Out of 1,810 patients, 1,674 to 1,291 patients depending upon time threshold, with 23 features were included for analysis and divided into training and validation sets. The best predictor had an average area under the curve (AUC) of 0.71 compared to existing clinical systems of 0.63 across all time points. On the heldout dataset, the predictor had an accuracy of around 0.74 at across all time thresholds with a specificity and sensitivity of 62% and 85% respectively. This machine learning approach was able to provide the best possible predictions of AVM radiosurgery outcomes of any method to date, identify a novel radiobiological feature (3D surface dose), and demonstrate a paradigm for further development of prognostic tools in medical care.

55Using a machine learning approach to predict outcome after surgery for degenerative cervical myelopathy.PubMed

Zamir G Merali, Christopher D Witiw, Jetan H Badhiwala, et al.
PLoS One. 2019 Apr 4;14(4):e0215133. doi: 10.1371/journal.pone.0215133. eCollection 2019.
Degenerative cervical myelopathy (DCM) is a spinal cord condition that results in progressive non-traumatic compression of the cervical spinal cord. Spine surgeons must consider a large quantity of information relating to disease presentation, imaging features, and patient characteristics to determine if a patient will benefit from surgery for DCM. We applied a supervised machine learning approach to develop a classification model to predict individual patient outcome after surgery for DCM. Patients undergoing surgery for DCM as a part of the AOSpine CSM-NA or CSM-I prospective, multi-centre studies were included in the analysis. Out of 757 patients 605, 583, and 539 patients had complete follow-up information at 6, 12, and 24 months respectively and were included in the analysis. The primary outcome was improvement in the SF-6D quality of life indicator score by the minimum clinically important difference (MCID). The secondary outcome was improvement in the modified Japanese Orthopedic Association (mJOA) score by the MCID. Predictor variables reflected information about pre-operative disease severity, disease presentation, patient demographics, and comorbidities. A machine learning approach of feature engineering, data pre-processing, and model optimization was used to create the most accurate predictive model of outcome after surgery for DCM. Following data pre-processing 48, 108, and 101 features were chosen for model training at 6, 12, and 24 months respectively. The best performing predictive model used a random forest structure and had an average area under the curve (AUC) of 0.70, classification accuracy of 77%, and sensitivity of 78% when evaluated on a testing cohort that was not used for model training. Worse pre-operative disease severity, longer duration of DCM symptoms, older age, higher body weight, and current smoking status were associated with worse surgical outcomes. We developed a model that predicted positive surgical outcome for DCM with good accuracy at the individual patient level on an independent testing cohort. Our analysis demonstrates the applicability of machine-learning to predictive modeling in spine surgery.

56Radiogenomics in Clear Cell Renal Cell Carcinoma: Machine Learning-Based High-Dimensional Quantitative CT Texture Analysis in Predicting PBRM1 Mutation Status.PubMed

Burak Kocak, Emine Sebnem Durmaz, Ece Ates, et al.
AJR Am J Roentgenol. 2019 Mar;212(3):W55-W63. doi: 10.2214/AJR.18.20443. Epub 2019 Jan 2.
OBJECTIVE: The purpose of this study is to evaluate the potential value of machine learning (ML)-based high-dimensional quantitative CT texture analysis in predicting the mutation status of the gene encoding the protein polybromo-1 (PBRM1) in patients with clear cell renal cell carcinoma (RCC). MATERIALS AND METHODS: In this retrospective study, 45 patients with clear cell RCC (29 without the PBRM1 mutation and 16 with the PBRM1 mutation) were identified in The Cancer Genome Atlas-Kidney Renal Clear Cell Carcinoma database. To create stable ML models and balanced classes, the data were augmented to a total of 161 labeled segmentations (87 without the PBRM1 mutation and 74 with the PBRM1 mutation) by obtaining three to five different samples per patient. Texture features were extracted from corticomedullary phase contrast-enhanced CT images with the use of an open-source software package for the extraction of radiomic data from medical images. Reproducibility analysis (intraclass correlation) was performed by two radiologists. Attribute selection and model optimization were done using a wrapper-based classifier-specific algorithm with nested cross-validation. ML classifiers were an artificial neural network (ANN) algorithm and a random forest (RF) algorithm. The models were validated using 10-fold cross-validation. The reference standard was the PBRM1 mutation status. The main performance metric was the AUC value. RESULTS: Of 828 extracted texture features, 759 had excellent reproducibility. Using 10 selected features, the ANN algorithm correctly classified 88.2% (142 of 161) of the clear cell RCCs in terms of PBRM1 mutation status (AUC value, 0.925). Using five selected features, the RF algorithm correctly classified 95.0% (153 of 161) of the clear cell RCCs (AUC value, 0.987). Overall, the RF algorithm performed better than the ANN algorithm (z score = -2.677; p = 0.007). CONCLUSION: ML-based high-dimensional quantitative CT texture analysis might be a feasible and potential method for predicting PBRM1 mutation status in patients with clear cell RCC.

57Machine learning algorithms for diagnosis of hip bone osteoporosis: a systematic review and meta-analysis study.PubMed

Fakher Rahim, Amin Zaki Zadeh, Pouya Javanmardi, et al.
Biomed Eng Online. 2023 Jul 10;22(1):68. doi: 10.1186/s12938-023-01132-9.
BACKGROUND: Osteoporosis is a significant health problem in the skeletal system, associated with bone tissue changes and its strength. Machine Learning (ML), on the other hand, has been accompanied by improvements in recent years and has been in the spotlight. This study is designed to investigate the Diagnostic Test Accuracy (DTA) of ML to detect osteoporosis through the hip dual-energy X-ray absorptiometry (DXA) images. METHODS: The ISI Web of Science, PubMed, Scopus, Cochrane Library, IEEE Xplore Digital Library, CINAHL, Science Direct, PROSPERO, and EMBASE were systematically searched until June 2023 for studies that tested the diagnostic precision of ML model-assisted for predicting an osteoporosis diagnosis. RESULTS: The pooled sensitivity of univariate analysis of seven studies was 0.844 (95% CI 0.791 to 0.885, I = 94% for 7 studies). The pooled specificity of univariate analysis was 0.781 (95% CI 0.732 to 0.824, I = 98% for 7 studies). The pooled diagnostic odds ratio (DOR) was 18.91 (95% CI 14.22 to 25.14, I = 93% for 7 studies). The pooled mean positive likelihood ratio (LR) and the negative likelihood ratio (LR) were 3.7 and 0.22, respectively. Also, the summary receiver operating characteristics (sROC) of the bivariate model has an AUC of 0.878. CONCLUSION: Osteoporosis can be diagnosed by ML with acceptable accuracy, and hip fracture prediction was improved via training in an Architecture Learning Network (ALN).

58Diagnostic accuracy of deep learning in prediction of osteoporosis: a systematic review and meta-analysis.PubMed

Firouz Amani, Masoud Amanzadeh, Mahnaz Hamedan, et al.
BMC Musculoskelet Disord. 2024 Dec 4;25(1):991. doi: 10.1186/s12891-024-08120-7.
BACKGROUND: Osteoporosis is one of the most common metabolic diseases that is characterized by a decrease in bone density and a loss of the quality of the bone structure. The use of deep learning in the prediction of osteoporosis can provide a non-invasive, cost-effective, and efficient approach. The aim of this study is to investigate the diagnostic accuracy of deep learning in the prediction of osteoporosis. METHODS: This is a systematic review and meta-analysis study that was conducted on the diagnostic accuracy of deep learning algorithms for predicting osteoporosis. A literature search was performed in electronic databases including PubMed, Elsevier, and Google Scholar to identify relevant articles until December 1, 2023. Articles were searched in databases by combining related terms such as "deep learning", "convolutional neural network", and "osteoporosis". We conducted title, abstract, and full-text screening based on inclusion/exclusion criteria. Various metrics, such as sensitivity, specificity, and area under the curve (AUC), were used to assess the diagnostic performance of deep learning models. RESULTS: Out of the 181 articles initially identified, 10 studies were included in the analysis. All studies used a convolutional neural network (CNN) as the deep learning model. Three studies investigated multiple deep learning models. Eight studies used various architectures of CNN, such as ResNet, VGG, and EfficientNet. The pooled sensitivity and specificity were 0.86 (95% CI, 0.82-0.89) and 0.89 (95% CI, 0.85-0.91), respectively. The bivariate approach's pooled SROC curve produced an AUC of 0.94 (95% CI 0.91-0.95). The Diagnostic Odds Ratio (DOR) for the deep learning models was 49.09 (95% CI, 28.74-83.84). Deeks' funnel plot asymmetry test (P = 0.4) suggested no potential publication bias. CONCLUSIONS: Deep learning has an acceptable performance for the diagnosis of osteoporosis, even better than other ML algorithms. However, further research is needed to validate the findings of this study in clinical trials.

59Diagnostic Test Accuracy of Deep Learning Detection of COVID-19: A Systematic Review and Meta-Analysis.PubMed

Temitope Emmanuel Komolafe, Yuzhu Cao, Benedictor Alexander Nguchu, et al.
Acad Radiol. 2021 Nov;28(11):1507-1523. doi: 10.1016/j.acra.2021.08.008. Epub 2021 Sep 17.
RATIONALE AND OBJECTIVE: To perform a meta-analysis to compare the diagnostic test accuracy (DTA) of deep learning (DL) in detecting coronavirus disease 2019 (COVID-19), and to investigate how network architecture and type of datasets affect DL performance. MATERIALS AND METHODS: We searched PubMed, Web of Science and Inspec from January 1, 2020, to December 3, 2020, for retrospective and prospective studies on deep learning detection with at least reported sensitivity and specificity. Pooled DTA was obtained using random-effect models. Sub-group analysis between studies was also carried out for data source and network architectures. RESULTS: The pooled sensitivity and specificity were 91% (95% confidence interval [CI]: 88%, 93%; I = 69%) and 92% (95% CI: 88%, 94%; I = 88%), respectively for 19 studies. The pooled AUC and diagnostic odds ratio (DOR) were 0.95 (95% CI: 0.88, 0.92) and 112.5 (95% CI: 57.7, 219.3; I = 90%) respectively. The overall accuracy, recall, F1-score, LR and LR are 89.5%, 89.5%, 89.7%, 23.13 and 0.13. Sub-group analysis shows that the sensitivity and DOR significantly vary with the type of network architectures and sources of data with low heterogeneity are (I = 0%) and (I = 18%) for ResNet architecture and single-source datasets, respectively. CONCLUSION: The diagnosis of COVID-19 via deep learning has achieved incredible performance, and the source of datasets, as well as network architectures, strongly affect DL performance.

60A Machine Learning Model for Prediction of Amputation in Diabetics.PubMed

Stavros Stefanopoulos, Qiong Qiu, Gang Ren, et al.
J Diabetes Sci Technol. 2024 Jul;18(4):874-881. doi: 10.1177/19322968221142899. Epub 2022 Dec 8.
BACKGROUND: Diabetic foot ulcer (DFU) and the resulting lower extremity amputation are associated with a poor survival prognosis. The objective of this study is to generate a model for predicting the probability of major amputation in hospitalized patients with DFU. METHODS: The National Inpatient Sample (NIS) database from 2008 to 2014 was used to select patients with DFU, who were then further divided by major amputation status. () and Agency for Healthcare Research and Quality (AHRQ) comorbidity codes were used to compare patient characteristics. For the descriptive statistics, the Student test, the χ test, and the Spearman correlation were utilized. The five most predictive variables were identified. A decision tree model (CTREE) based on conditional inference framework algorithm and a random forest model were used to develop the algorithm. RESULTS: A total of 326 853 inpatients with DFU were identified, and 5.9% underwent major amputation. The top five contributory variables (all with < .001) were gangrene (odds ratio [OR] = 11.8, 95% confidence interval [CI] = 11.5-12.2), peripheral vascular disease (OR = 2.9, 95% CI = 2.8-3.0), weight loss (OR = 2.6, 95% CI = 2.5-2.8), systemic infection (OR = 2.5, 95% CI = 2.4-2.53), and osteomyelitis (OR = 1.7, 95% CI = 1.6-1.73). The model performance of the training data was 77.7% (76.1% sensitivity and 79.3% specificity) and of the testing data was 77.8% (76.2% sensitivity and 79.4% specificity). The model was further validated with boosting and random forest models which demonstrated similar performance and area under the curve (AUC) (0.84, 95% CI = 0.83-0.85). CONCLUSION: Utilizing machine learning methods, we have developed a clinical algorithm that predicts the risk of major lower extremity amputation for inpatients with diabetes with 77.8% accuracy.

61Management of Lung Cancer in the Patient with Interstitial Lung Disease.PubMed

Angela J Frank, Ibiayi Dagogo-Jack, Ioana A Dobre, et al.
Oncologist. 2023 Jan 18;28(1):12-22. doi: 10.1093/oncolo/oyac226.
Patients with interstitial lung disease (ILD), especially those with pulmonary fibrosis, are at increased risk of developing lung cancer. Management of lung cancer in patients with ILD is particularly challenging. Diagnosis can be complicated by difficulty differentiating lung nodules from areas of focal fibrosis, and percutaneous biopsy approaches confer an increased risk of complications in those with pulmonary fibrosis. Lung cancer treatment in these patients pose several specific considerations. The degree of lung function impairment may preclude lobectomy or surgical resection of any type. Surgical resection can trigger an acute exacerbation of the underlying ILD. The presence of ILD confers an increased risk of pneumonitis with radiotherapy, and many of the systemic therapies also carry an increased risk of pneumonitis in this population. The safety of immunotherapy in the setting of ILD remains to be fully elucidated and concerns remain as to triggering pneumonitis. The purpose of this review is to summarize the evidence regarding consideration for tissue diagnosis, chemotherapy and immunotherapy, radiotherapy, and surgery, in this patient population and discuss emerging areas of research. We also propose a multidisciplinary approach and practical considerations for monitoring for ILD progression during lung cancer treatment.

62Risk factors and immunomodulators use in steroid-refractory checkpoint inhibitor pneumonitis.PubMed

Yanlin Li, Xiaohui Jia, Yajuan Zhang, et al.
J Immunother Cancer. 2023 Jun;11(6). doi: 10.1136/jitc-2023-006982.
BACKGROUND: Checkpoint inhibitor pneumonitis (CIP) that does not respond to corticosteroids is termed steroid-refractory CIP. We aimed to find risk factors of steroid-refractory CIP and evaluate the management strategies of immunomodulators (IMs). METHODS: Patients with CIP were identified between August 2019 and August 2022 retrospectively. Clinical characteristics, peripheral blood biomarkers, and radiologic images were collected. RESULTS: Among 1209 patients with solid tumor receiving programmed death (ligand)-1 antibody, 28 patients developed steroid-refractory CIP and 38 patients developed steroid-response CIP. Patients with steroid-refractory CIP had a higher proportion of previous interstitial lung disease (p=0.015) and grade 3-4 (p<0.001) at diagnosis. Otherwise, absolute neutrophil count (ANC), procalcitonin were higher and albumin was lower in steroid-refractory patients (ANC, p=0.009; procalcitonin, p=0.024; albumin, p=0.026). After multivariate analysis, grade 3-4 and higher ANC at diagnosis were confirmed to be independent risk factors for steroid-refractory CIP (grade, p=0.001; ANC, p=0.046). For grade 2 steroid-refractory CIP, additional IMs did not affect the prognosis (p=1.000). However, additional IMs reduced the risk of deterioration significantly in grade 3-4 steroid-refractory CIP (p=0.036). CONCLUSIONS: Grade 3-4 and higher peripheral blood ANC at diagnosis are associated with higher risk of steroid-refractory CIP. The use of additional IMs improves the outcome of grade 3-4 steroid-refractory CIP. These results can offer new insights to the decision-making of CIP management.

63Radiation Versus Immune Checkpoint Inhibitor Associated Pneumonitis: Distinct Radiologic Morphologies.PubMed

Xuguang Chen, Khadija Sheikh, Erica Nakajima, et al.
Oncologist. 2021 Oct;26(10):e1822-e1832. doi: 10.1002/onco.13900. Epub 2021 Aug 4.
BACKGROUND: Patients with non-small cell lung cancer may develop pneumonitis after thoracic radiotherapy (RT) and immune checkpoint inhibitors (ICIs). We hypothesized that distinct morphologic features are associated with different pneumonitis etiologies. MATERIALS AND METHODS: We systematically compared computed tomography (CT) features of RT- versus ICI-pneumonitis. Clinical and imaging features were tested for association with pneumonitis severity. Lastly, we constructed an exploratory radiomics-based machine learning (ML) model to discern pneumonitis etiology. RESULTS: Between 2009 and 2019, 82 patients developed pneumonitis: 29 after thoracic RT, 23 after ICI, and 30 after RT + ICI. Fifty patients had grade 2 pneumonitis, 22 grade 3, and 7 grade 4. ICI-pneumonitis was more likely bilateral (65% vs. 28%; p = .01) and involved more lobes (66% vs. 45% involving at least three lobes) and was less likely to have sharp border (17% vs. 59%; p = .004) compared with RT-pneumonitis. Pneumonitis morphology after RT + ICI was heterogeneous, with 47% bilateral, 37% involving at least three lobes, and 40% sharp borders. Among all patients, risk factors for severe pneumonitis included poor performance status, smoking history, worse lung function, and bilateral and multifocal involvement on CT. An ML model based on seven radiomic features alone could distinguish ICI- from RT-pneumonitis with an area under the receiver-operating curve of 0.76 and identified the predominant etiology after RT + ICI concordant with multidisciplinary consensus. CONCLUSION: RT- and ICI-pneumonitis exhibit distinct spatial features on CT. Bilateral and multifocal lung involvement is associated with severe pneumonitis. Integrating these morphologic features in the clinical management of patients who develop pneumonitis after RT and ICIs may improve treatment decision-making. IMPLICATIONS FOR PRACTICE: Patients with non-small cell lung cancer often receive thoracic radiation and immune checkpoint inhibitors (ICIs), both of which can cause pneumonitis. This study identified similarities and differences in pneumonitis morphology on computed tomography (CT) scans among pneumonitis due to radiotherapy (RT) alone, ICI alone, and the combination of both. Patients who have bilateral CT changes involving at least three lobes are more likely to have ICI-pneumonitis, whereas those with unilateral CT changes with sharp borders are more likely to have radiation pneumonitis. After RT and/or ICI, severe pneumonitis is associated with bilateral and multifocal CT changes. These results can help guide clinicians in triaging patients who develop pneumonitis after radiation and during ICI treatment.

64Validated machine learning tools to distinguish immune checkpoint inhibitor, radiotherapy, COVID-19 and other infective pneumonitis.PubMed

Sumeet Hindocha, Benjamin Hunter, Kristofer Linton-Reid, et al.
Radiother Oncol. 2024 Jun;195:110266. doi: 10.1016/j.radonc.2024.110266. Epub 2024 Apr 4.
BACKGROUND: Pneumonitis is a well-described, potentially disabling, or fatal adverse effect associated with both immune checkpoint inhibitors (ICI) and thoracic radiotherapy. Accurate differentiation between checkpoint inhibitor pneumonitis (CIP) radiation pneumonitis (RP), and infective pneumonitis (IP) is crucial for swift, appropriate, and tailored management to achieve optimal patient outcomes. However, correct diagnosis is often challenging, owing to overlapping clinical presentations and radiological patterns. METHODS: In this multi-centre study of 455 patients, we used machine learning with radiomic features extracted from chest CT imaging to develop and validate five models to distinguish CIP and RP from COVID-19, non-COVID-19 infective pneumonitis, and each other. Model performance was compared to that of two radiologists. RESULTS: Models to distinguish RP from COVID-19, CIP from COVID-19 and CIP from non-COVID-19 IP out-performed radiologists (test set AUCs of 0.92 vs 0.8 and 0.8; 0.68 vs 0.43 and 0.4; 0.71 vs 0.55 and 0.63 respectively). Models to distinguish RP from non-COVID-19 IP and CIP from RP were not superior to radiologists but demonstrated modest performance, with test set AUCs of 0.81 and 0.8 respectively. The CIP vs RP model performed less well on patients with prior exposure to both ICI and radiotherapy (AUC 0.54), though the radiologists also had difficulty distinguishing this test cohort (AUC values 0.6 and 0.6). CONCLUSION: Our results demonstrate the potential utility of such tools as a second or concurrent reader to support oncologists, radiologists, and chest physicians in cases of diagnostic uncertainty. Further research is required for patients with exposure to both ICI and thoracic radiotherapy.

65Performance and clinical utility of an artificial intelligence-enabled tool for pulmonary embolism detection.PubMed

Angela Ayobi, Peter D Chang, Daniel S Chow, et al.
Clin Imaging. 2024 Sep;113:110245. doi: 10.1016/j.clinimag.2024.110245. Epub 2024 Jul 30.
PURPOSE: Diagnosing pulmonary embolism (PE) is still challenging due to other conditions that can mimic its appearance, leading to incomplete or delayed management and several inter-observer variabilities. This study evaluated the performance and clinical utility of an artificial intelligence (AI)-based application designed to assist clinicians in the detection of PE on CT pulmonary angiography (CTPA). PATIENTS AND METHODS: CTPAs from 230 US cities acquired on 57 scanner models from 6 different vendors were retrospectively collected. Three US board certified expert radiologists defined the ground truth by majority agreement. The same cases were analyzed by CINA-PE, an AI-driven algorithm capable of detecting and highlighting suspected PE locations. The algorithm's performance at a per-case and per-finding level was evaluated. Furthermore, cases with PE not mentioned in the clinical report but correctly detected by the algorithm were analyzed. RESULTS: A total of 1204 CTPAs (mean age 62.1 years ± 16.6[SD], 44.4 % female, 14.9 % positive) were included in the study. Per-case sensitivity and specificity were 93.9 % (95%CI: 89.3 %-96.9 %) and 94.8 % (95%CI: 93.3 %-96.1 %), respectively. Per-finding positive predictive value was 89.5 % (95%CI: 86.7 %-91.9 %). Among the 196 positive cases, 29 (15.6 %) were not mentioned in the clinical report. The algorithm detected 22/29 (76 %) of these cases, leading to a reduction in the miss rate from 15.6 % to 3.8 % (7/186). CONCLUSIONS: The AI-based application may improve diagnostic accuracy in detecting PE and enhance patient outcomes through timely intervention. Integrating AI tools in clinical workflows can reduce missed or delayed diagnoses, and positively impact healthcare delivery and patient care.

66Development and Validation of a Machine Learning-Based Model Using CT Radiomics for Predicting Immune Checkpoint Inhibitor-related Pneumonitis in Patients With NSCLC Receiving Anti-PD1 Immunotherapy: A Multicenter Retrospective CaseControl Study.PubMed

Guo-Yue Zhang, Xian-Zhi Du, Rui Xu, et al.
Acad Radiol. 2024 May;31(5):2128-2143. doi: 10.1016/j.acra.2023.10.039. Epub 2023 Nov 17.
RATIONALE AND OBJECTIVES: This study aimed to develop and evaluate a radiomics-based model combined with clinical and qualitative radiological (semantic feature [SF]) features to predict immune checkpoint inhibitor-related pneumonitis (CIP) in patients with non-small cell lung cancer (NSCLC) treated with programmed cell death protein 1 inhibitors. MATERIALS AND METHODS: This was a multicenter retrospective casecontrol study conducted from January 1, 2018, to December 31, 2022, at three centers. Patients with NSCLC treated with anti-PD1 were enrolled and randomly divided into two groups (7:3): training (n = 95) and validation (n = 39). Logistic regression (LR) and support vector machine (SVM) algorithms were used to transform features into the models. RESULTS: The study comprised 134 participants from three independent centers (male, 114/134, 85%; mean [±standard deviation] age, 63.92 [±7.9] years). The radiomics score (RS) models built based on the LR and SVM algorithms could accurately predict CIP (area under the receiver operating characteristics curve [AUC], 0.860 [0.780, 0.939] and 0.861 [0.781, 0.941], respectively). The AUCs for the RS-clinic-SF combined model were 0.903 (0.839, 0.967) and 0.826 (0.688, 0.964) in the training and validation cohorts, respectively. Decision curve analysis showed that the combined models achieved high clinical net benefit across the majority of the range of reasonable threshold probabilities. CONCLUSION: This study demonstrated that the combined model constructed by the identified features of RS, clinical features, and SF has the potential to precisely predict CIP. The RS-clinic-SF combined model has the potential to be used more widely as a practical tool for the noninvasive prediction of CIP to support individualized treatment planning.

67Identification and prediction of immune checkpoint inhibitors-related pneumonitis by machine learning.PubMed

Li Gong, Jun Gong, Xin Sun, et al.
Front Immunol. 2023 Jun 29;14:1138489. doi: 10.3389/fimmu.2023.1138489. eCollection 2023.
BACKGROUND: Immune checkpoint inhibitor (ICI)-related pneumonitis (IRP) is a common and potentially fatal clinical adverse event. The identification and prediction of the risk of ICI-related IRP is a major clinical issue. The objective of this study was to apply a machine learning method to explore risk factors and establish a prediction model. METHODS: We retrospectively analyzed 48 patients with IRP (IRP group) and 142 patients without IRP (control group) who were treated with ICIs. An Elastic Net model was constructed using a repeated k-fold cross-validation framework (repeat = 10; k = 3). The prediction models were validated internally and the final prediction model was built on the entire training set using hyperparameters with the best interval validation performance. The generalizability of the final prediction model was assessed by applying it to an independent test set. The overall performance, discrimination, and calibration of the prediction model were evaluated. RESULTS: Eleven predictors were included in the final predictive model: sindillizumab, number of ≥2 underlying diseases, history of lung diseases, tirelizumab, non-small cell lung cancer (NSCLC), percentage of CD4 lymphocytes, body temperature, KPS score ≤70, hemoglobin, cancer stage IV, and history of antitumor therapy. The external validation of the risk prediction model on an independent test set of 37 patients and showed good discrimination and acceptable calibration ability: with AUC of 0.81 (95% CI 0.58-0.90), AP of 0.76, scaled Brier score of 0.31, and Spiegelhalter-z of -0.29 (P-value:0.77). We also designed an online IRP risk calculator for use in clinical practice. CONCLUSION: The prediction model of ICI-related IRP provides a tool for accurately predicting the occurrence of IRP in patients with cancer who received ICIs.

68Pushing the Boundaries of Molecular Representation for Drug Discovery with the Graph Attention Mechanism.PubMed

Zhaoping Xiong, Dingyan Wang, Xiaohong Liu, et al.
J Med Chem. 2020 Aug 27;63(16):8749-8760. doi: 10.1021/acs.jmedchem.9b00959. Epub 2019 Aug 27.
Hunting for chemicals with favorable pharmacological, toxicological, and pharmacokinetic properties remains a formidable challenge for drug discovery. Deep learning provides us with powerful tools to build predictive models that are appropriate for the rising amounts of data, but the gap between what these neural networks learn and what human beings can comprehend is growing. Moreover, this gap may induce distrust and restrict deep learning applications in practice. Here, we introduce a new graph neural network architecture called Attentive FP for molecular representation that uses a graph attention mechanism to learn from relevant drug discovery data sets. We demonstrate that Attentive FP achieves state-of-the-art predictive performances on a variety of data sets and that what it learns is interpretable. The feature visualization for Attentive FP suggests that it automatically learns nonlocal intramolecular interactions from specified tasks, which can help us gain chemical insights directly from data beyond human perception.

69Predicting alcohol dependence from multi-site brain structural measures.PubMed

Sage Hahn, Scott Mackey, Janna Cousijn, et al.
Hum Brain Mapp. 2022 Jan;43(1):555-565. doi: 10.1002/hbm.25248. Epub 2020 Oct 16.
To identify neuroimaging biomarkers of alcohol dependence (AD) from structural magnetic resonance imaging, it may be useful to develop classification models that are explicitly generalizable to unseen sites and populations. This problem was explored in a mega-analysis of previously published datasets from 2,034 AD and comparison participants spanning 27 sites curated by the ENIGMA Addiction Working Group. Data were grouped into a training set used for internal validation including 1,652 participants (692 AD, 24 sites), and a test set used for external validation with 382 participants (146 AD, 3 sites). An exploratory data analysis was first conducted, followed by an evolutionary search based feature selection to site generalizable and high performing subsets of brain measurements. Exploratory data analysis revealed that inclusion of case- and control-only sites led to the inadvertent learning of site-effects. Cross validation methods that do not properly account for site can drastically overestimate results. Evolutionary-based feature selection leveraging leave-one-site-out cross-validation, to combat unintentional learning, identified cortical thickness in the left superior frontal gyrus and right lateral orbitofrontal cortex, cortical surface area in the right transverse temporal gyrus, and left putamen volume as final features. Ridge regression restricted to these features yielded a test-set area under the receiver operating characteristic curve of 0.768. These findings evaluate strategies for handling multi-site data with varied underlying class distributions and identify potential biomarkers for individuals with current AD.

70Lung Inflammation Predictors in Combined Immune Checkpoint-Inhibitor and Radiation Therapy-Proof-of-Concept Animal Study.PubMed

Benjamin Spieler, Teresa M Giret, Scott Welford, et al.
Biomedicines. 2022 May 19;10(5):1173. doi: 10.3390/biomedicines10051173.
PURPOSE: Combined radiotherapy (RT) and immune checkpoint-inhibitor (ICI) therapy can act synergistically to enhance tumor response beyond what either treatment can achieve alone. Alongside the revolutionary impact of ICIs on cancer therapy, life-threatening potential side effects, such as checkpoint-inhibitor-induced (CIP) pneumonitis, remain underreported and unpredictable. In this preclinical study, we hypothesized that routinely collected data such as imaging, blood counts, and blood cytokine levels can be utilized to build a model that predicts lung inflammation associated with combined RT/ICI therapy. MATERIALS AND METHODS: This proof-of-concept investigational work was performed on Lewis lung carcinoma in a syngeneic murine model. Nineteen mice were used, four as untreated controls and the rest subjected to RT/ICI therapy. Tumors were implanted subcutaneously in both flanks and upon reaching volumes of ~200 mm the animals were imaged with both CT and MRI and blood was collected. Quantitative radiomics features were extracted from imaging of both lungs. The animals then received RT to the right flank tumor only with a regimen of three 8 Gy fractions (one fraction per day over 3 days) with PD-1 inhibitor administration delivered intraperitoneally after each daily RT fraction. Tumor volume evolution was followed until tumors reached the maximum size allowed by the Institutional Animal Care and Use Committee (IACUC). The animals were sacrificed, and lung tissues harvested for immunohistochemistry evaluation. Tissue biomarkers of lung inflammation (CD45) were tallied, and binary logistic regression analyses were performed to create models predictive of lung inflammation, incorporating pretreatment CT/MRI radiomics, blood counts, and blood cytokines. RESULTS: The treated animal cohort was dichotomized by the median value of CD45 infiltration in the lungs. Four pretreatment radiomics features (3 CT features and 1 MRI feature) together with pre-treatment neutrophil-to-lymphocyte (NLR) ratio and pre-treatment granulocyte-macrophage colony-stimulating factor (GM-CSF) level correlated with dichotomized CD45 infiltration. Predictive models were created by combining radiomics with NLR and GM-CSF. Receiver operating characteristic (ROC) analyses of two-fold internal cross-validation indicated that the predictive model incorporating MR radiomics had an average area under the curve (AUC) of 0.834, while the model incorporating CT radiomics had an AUC of 0.787. CONCLUSIONS: Model building using quantitative imaging data, blood counts, and blood cytokines resulted in lung inflammation prediction models justifying the study hypothesis. The models yielded very-good-to-excellent AUCs of more than 0.78 on internal cross-validation analyses.

71Immune Biomarkers for Checkpoint Blockade in Solid Tumors: Transitioning from Tissue to Peripheral Blood Monitoring and Future Integrated Strategies.PubMed

Ioannis P Trontzas, Konstantinos N Syrigos
Cancers (Basel). 2025 Aug 13;17(16):2639. doi: 10.3390/cancers17162639.
Immunotherapy with immune checkpoint inhibitors has changed the treatment landscape in many solid tumors. Despite the unprecedent success, many patients will develop primary or secondary resistance to treatment or will hold up therapy due to the emerging immune-related toxicity. Traditionally, tissue-based immune biomarkers, such as PD-L1 expression, have been used to select patients who will benefit most from immunotherapy. However, these markers demonstrate major limitations, such as tumor heterogeneity and sample constraints. In addition, they do not reflect the dynamic interplay of tumor and hosts immune response during treatment. Peripheral blood immunomarkers offer a minimally invasive, real-time assessment of the immune system and its interaction with the tumor. Integration of traditional tissue-based and peripheral blood markers coupled with the recent developments in computational platforms, artificial intelligence, and machine learning models may provide more successful biomarkers for prognosis, prediction of immunotherapy-related outcomes, the early evaluation of forthcoming disease progression, and the prediction of the emerging immune-related adverse events. Despite the promising developments in the field of immune biomarkers, several issues including assay standardization, clinical validation, and biological variability should be addressed to improve personalized immunotherapy approaches. In this comprehensive review we provide an update on immune biomarker evolution, and we discuss the current limitations and future directions.