Chen Yanqiu, Sun Zhen, Zhong Huohu, Chen Yuwei, Wu Xiuming, Su Liyang, Lai Zhenhan, Zheng Tao, Lyu Guorong, Su Qichen
Department of Ultrasound, the Second Affiliated Hospital of Fujian Medical University, Fujian, China.
Department of Clinical Medicine, Quanzhou Medical College, Fujian, China.
Ultrasound Med Biol. 2025 Jul 12. doi: 10.1016/j.ultrasmedbio.2025.06.016.
This study aimed to develop and evaluate eight machine learning models based on multimodal ultrasound to precisely predict of diabetic tibial neuropathy (DTN) in patients. Additionally, the SHapley Additive exPlanations(SHAP)framework was introduced to quantify the importance of each feature variable, providing a precise and noninvasive assessment tool for DTN patients, optimizing clinical management strategies, and enhancing patient prognosis.
A prospective analysis was conducted using multimodal ultrasound and clinical data from 255 suspected DTN patients who visited the Second Affiliated Hospital of Fujian Medical University between January 2024 and November 2024. Key features were selected using Least Absolute Shrinkage and Selection Operator (LASSO) regression. Predictive models were constructed using Extreme Gradient Boosting (XGB), Logistic Regression, Support Vector Machines, k-Nearest Neighbors, Random Forest, Decision Tree, Naïve Bayes, and Neural Network. The SHAP method was employed to refine model interpretability. Furthermore, in order to verify the generalization degree of the model, this study also collected 135 patients from three other tertiary hospitals for external test.
LASSO regression identified Echo intensity(EI), Cross-sectional area (CSA), Mean elasticity value(Emean), Superb microvascular imaging(SMI), and History of smoking were key features for DTN prediction. The XGB model achieved an Area Under the Curve (AUC) of 0.94, 0.83 and 0.79 in the training, internal test and external test sets, respectively. SHAP analysis highlighted the ranking significance of EI, CSA, Emean, SMI, and History of smoking. Personalized prediction explanations provided by theSHAP values demonstrated the contribution of each feature to the final prediction, and enhancing model interpretability. Furthermore, decision plots depicted how different features influenced mispredictions, thereby facilitating further model optimization or feature adjustment.
This study proposed a DTN prediction model based on machine-learning algorithms applied to multimodal ultrasound data. The results indicated the superior performance of the XGB model and its interpretability was enhanced using SHAP analysis. This cost-effective and user-friendly approach provides potential support for personalized treatment and precision medicine for DTN.
本研究旨在开发并评估基于多模态超声的八种机器学习模型，以精确预测患者的糖尿病性胫神经病变（DTN）。此外，引入了SHapley值相加解释（SHAP）框架来量化每个特征变量的重要性，为DTN患者提供一种精确且无创的评估工具，优化临床管理策略，并改善患者预后。
对2024年1月至2024年11月期间就诊于福建医科大学附属第二医院的255例疑似DTN患者的多模态超声和临床数据进行前瞻性分析。使用最小绝对收缩和选择算子（LASSO）回归选择关键特征。采用极端梯度提升（XGB）、逻辑回归、支持向量机、k近邻、随机森林、决策树、朴素贝叶斯和神经网络构建预测模型。采用SHAP方法提高模型的可解释性。此外，为了验证模型的泛化程度，本研究还收集了来自其他三家三级医院的135例患者进行外部测试。
LASSO回归确定回声强度（EI）、横截面积（CSA）、平均弹性值（Emean）、超微血管成像（SMI）和吸烟史是DTN预测的关键特征。XGB模型在训练集、内部测试集和外部测试集中的曲线下面积（AUC）分别为0.94、0.83和0.79。SHAP分析突出了EI、CSA、Emean、SMI和吸烟史的排名重要性。SHAP值提供的个性化预测解释展示了每个特征对最终预测的贡献，增强了模型的可解释性。此外，决策图描绘了不同特征如何影响错误预测，从而有助于进一步优化模型或调整特征。
本研究提出了一种基于应用于多模态超声数据的机器学习算法的DTN预测模型。结果表明XGB模型具有卓越性能，且通过SHAP分析增强了其可解释性。这种经济高效且用户友好的方法为DTN的个性化治疗和精准医学提供了潜在支持。
