AutoCOPD——一种使用全肺吸气定量CT测量进行慢性阻塞性肺疾病（COPD）检测的新型实用机器学习模型：一项回顾性多中心研究

AutoCOPD-A novel and practical machine learning model for COPD detection using whole-lung inspiratory quantitative CT measurements: a retrospective, multicenter study.

作者信息

Lin Fanjie, Zhang Zili, Wang Jian, Liang Cuixia, Xu Jiaxuan, Zeng Xiansheng, Zeng Qingpeng, Chen Huai, Zhuang Jiayu, Ma Yu, Ma Qiao, Shi Raymond, Xu Jingyi, Li Yuanyuan, Yuan Liang, Wei Xinguang, Wu Lulu, Huang Renjun, Xiao Tianchi, Liang Wenhua, Zheng Jinping, He Jianxing, Liu Yun, Liang Zhenyu, Zhong Nanshan, Lu Wenju

机构信息

State Key Laboratory of Respiratory Disease, Guangdong Key Laboratory of Vascular Disease, National Clinical Research Center for Respiratory Disease, National Center for Respiratory Medicine, Guangzhou Institute of Respiratory Health, The First Affiliated Hospital of Guangzhou Medical University, Guangzhou, Guangdong, PR China.

Guangzhou National Lab, Guangzhou, Guangdong, PR China.

出版信息

EClinicalMedicine. 2025 Apr 3;82:103166. doi: 10.1016/j.eclinm.2025.103166. eCollection 2025 Apr.

DOI:10.1016/j.eclinm.2025.103166

PMID:40242563

原文链接:https://pmc.ncbi.nlm.nih.gov/articles/PMC12002883/

Abstract

BACKGROUND

The rate of diagnosis for chronic obstructive pulmonary disease (COPD) is low worldwide. Quantitative computed tomography (QCT) parameters add value to quantify alterations in airway and lung parenchyma for COPD. This study aimed to assess the performance of QCT features in COPD detection using a whole-lung inspiratory CT model.

METHODS

This multicenter retrospective study was performed on 4106 participants. The derivation cohort containing 1950 participants who enrolled in Guangzhou communities from August 2017 to December 2019, was separated for training and internal validation cohorts, and three external validation cohorts containing 1703 participants were recruited from the public hospitals (Cohort 1: the First Affiliated Hospital of Guangzhou Medical University; Cohort 2: Xiangyang central hospital; Cohort 3: the Second Affiliated Hospital of Xi'an Jiaotong University) in China between April 2017 and May 2024. Questionnaire information, CT reports, and QCT features derived from inspiratory CT were extracted for model development. A novel multimodal framework using eXtreme gradient boosting and hybrid feature selection was established for COPD detection. National Lung Screening Trial (NLST) cohort (n = 453) was applied to validate the multiracial extrapolation and robustness on low-dose CT scans.

FINDINGS

The QCT model (referred to as AutoCOPD) with ten features achieved the highest AUC of 0·860 (95% CI: 0·823-0·898) in the internal validation cohort, and showed excellent discrimination when externally validated [Cohort 1: AUC = 0·915 (95% CI: 0·898-0·931); Cohort 2: AUC = 0·903 (95% CI: 0·864-0·943); Cohort 3: AUC = 0·914 (95% CI: 0·882-0·947); NLST: AUC = 0·881 (95% CI: 0·846-0·915)]. Decision curve analysis demonstrated that AutoCOPD was valuable across a range of COPD risk thresholds between 0·12 and 0·66 compared with intervention in all patients with COPD or no intervention.

INTERPRETATION

Heterogeneous COPD can be well identified using AutoCOPD (https://lwj-lab.shinyapps.io/autocopd/) constructed by a subset of only ten QCT features. It may be generalizable across clinical settings and serve as a feasible tool for early detecting patients with mild or asymptomatic COPD to reduce delayed diagnosis in routine practice.

FUNDING

The National Natural Science Foundation of China, Guangzhou Laboratory, Natural Science Foundation of Guangdong Province, Guangzhou Municipal Science and Technology grant, State Key Laboratory of Respiratory Disease.

摘要

背景

慢性阻塞性肺疾病（COPD）在全球范围内的诊断率较低。定量计算机断层扫描（QCT）参数有助于量化COPD患者气道和肺实质的改变。本研究旨在使用全肺吸气CT模型评估QCT特征在COPD检测中的性能。

方法

对4106名参与者进行了这项多中心回顾性研究。将2017年8月至2019年12月在广州社区招募的1950名参与者组成的推导队列分为训练队列和内部验证队列，并于2017年4月至2024年5月在中国的公立医院招募了三个包含1703名参与者的外部验证队列（队列1：广州医科大学附属第一医院；队列2：襄阳市中心医院；队列3：西安交通大学第二附属医院）。提取问卷信息、CT报告以及吸气CT得出的QCT特征用于模型开发。建立了一种使用极端梯度提升和混合特征选择的新型多模态框架用于COPD检测。应用国家肺癌筛查试验（NLST）队列（n = 453）验证低剂量CT扫描的多种族外推和稳健性。

结果

具有十个特征的QCT模型（称为AutoCOPD）在内部验证队列中达到了最高的AUC为0.860（95% CI：0.823 - 0.898），在外部验证时表现出优异的区分能力[队列1：AUC = 0.915（95% CI：0.898 - 0.931）；队列2：AUC = 0.903（95% CI：0.864 - 0.943）；队列3：AUC = 0.914（95% CI：0.882 - 0.947）；NLST：AUC = 0.881（95% CI：0.846 - 0.915）]。决策曲线分析表明，与对所有COPD患者进行干预或不进行干预相比，AutoCOPD在0.12至0.66的一系列COPD风险阈值范围内都具有价值。