基于肺参数响应图和机器学习对高危慢性阻塞性肺疾病肺功能诊断标准的初步研究  

Preliminary study on diagnostic criteria of lung function based on lung parameter response mapping and machine learning for high-risk chronic obstructive pulmonary disease

作　　者：周秀秀 蒲瑜 张迪 管宇 夏艺 涂文婷 刘士远 范丽 ZHOU Xiuxiu;PU Yu;ZHANG Di;GUAN Yu;XIA Yi;TU Wenting;LIU Shiyuan;FAN Li(Department of Diagnostic Radiolog y,the Second Affiliated Hospital of PLA Naval Military Medical University,Shanghai 200003,China)

机构地区：[1]中国人民解放军海军军医大学第二附属医院放射诊断科,上海200003

出　　处：《实用放射学杂志》2023年第5期714-718,共5页Journal of Practical Radiology

基　　金：国家自然科学基金项目(81871321,81930049,82171926);上海市科学技术委员会计划项目(21DZ2202600);上海市青年科技英才扬帆计划项目(20YF1449000)。

摘　　要：目的基于肺参数响应图(PRM)的机器学习分类模型,初步探讨诊断慢性阻塞性肺疾病(COPD)高危患者的第1 s用力呼气量占预计值百分比(FEV1%)最佳阈值。方法回顾性分析进行胸部疾病筛查的615例受试者,依据肺功能检测(PFT)分为非COPD组和COPD组,非COPD组根据不同的FEV1%阈值分为正常组和高危组。通过双呼吸相CT图像,分别得到全肺、双肺及5个肺叶水平的72个PRM定量参数。分别分析年龄、体质量指数(BMI)及PRM参数的组间差异;将FEV1%阈值从50%~129%,以1%为间隔,划分为80个阈值,在每个阈值下建立一个基于PRM的正常和高危的随机森林分类模型,并进行诊断效果分析。结果年龄、BMI在非COPD组和COPD组间差异无统计学意义(P>0.05);PRM参数中的肺体积(PRM^(LV))、肺气肿体积(PRM^(EmphV))、功能小气道疾病体积(PRM^(fSADV))及其百分比在非COPD组和COPD组间有统计学差异(P<0.05),并且非COPD组中PRM^(EmphV)、PRM^(fSADV)和PRM^(LV)均低于COPD组(P<0.05)。当FEV1%=72%时,PRM参数和PFT在区分正常和高危COPD的一致性较好,平均曲线下面积(AUC)为0.83;当FEV1%阈值取95%及80%时,平均AUC分别为0.64、0.72。结论基于PRM参数的机器学习模型用于诊断高危COPD是可行的,推荐FEV1%阈值取72%作为高危COPD的PFT诊断标准。Objective To explore the optimal threshold of the percentage of forced expiratory volume in one second predicted(FEV1%)in high-risk chronic obstructive pulmonary disease(COPD)patients by machine learning classification model based on lung parameter response mapping(PRM).Methods A total of 615 subjects screened from chest diseases were analyzed retrospectively.They were classified into non-COPD group and COPD group according to the results of pulmonary function test(PFT).The non-COPD group was divided into normal group and high-risk group according to different FEV1%thresholds.The 72 quantitative parameters of PRM at the level of whole lung,two lungs and five lobes were obtained by bierspiratory CT images.The differences of age,body mass index(BMI)and PRM parameters were analyzed.The FEV1%threshold was devided into 80 thresholds from 50%to 129%with an interval of 1%.The normal and high-risk random forest classification model under each threshold based on PRM was established,and the performance of each model was analyzed.Results There were no significant differences in age and BMI between the non-COPD group and COPD group(P>0.05).There were significant differences in PRM^(LV),PRM^(EmphV),PRM^(fSADV) and their corresponding percentages between non-COPD group and COPD group,and PRM^(LV),PRM^(EmphV) and PRM^(fSADV) in non-COPD group were lower than those in COPD group(P<0.05).When the FEV1%was 72%,the consistency between PRM parameters and PFT in distinguishing high-risk COPD from normal was good,and the average area under the curve(AUC)was 0.83.When the FEV1%threshold was set as 80%or 95%,the average AUC was 0.72 or 0.64,respectively.Conclusion The machine learning model based on PRM parameters is feasible to diagnosis high-risk COPD,and the optimal FEV1%of 72%is recommended as the PFT diagnostic criteria of high-risk COPD.

关 键 词：慢性阻塞性肺疾病 功能小气道疾病 定量影像 肺功能检测 机器学习 计算机体层成像 

分 类 号：R563[医药卫生—呼吸系统] TP181[医药卫生—内科学] R814.42[医药卫生—临床医学]