Predicting Acute Exacerbation Phenotype in Chronic Obstructive Pulmonary Disease Patients using VGG-16 Deep Learning.

Abstract

We performed a retrospective study on 219 patients with COPD who underwent inspiratory and expiratory HRCT scans. By recording the acute respiratory events of the previous year, these patients were further divided into Non-AECOPD group and AECOPD group according to the presence of acute exacerbation events. 69 Quantitative CT (QCT) parameters of emphysema and airway were calculated by NeuLungCARE software and 2000 DL features were extracted by VGG-16 method. The Logistic regression method was employed to identify AECOPD patients and 29 patients of external validation cohort were used to access the robustness of the results. The Model 3-B achieved an AUC of 0.933, and 0.865 in the testing cohort and external validation cohort respectively. Model 3-I obtained AUC of 0.895 in the testing cohort and AUC of 0.774 in the external validation cohort. Model 7-B combined clinical characteristics, QCT parameters, and DL features achieved the best performance with an AUC of 0.979 in the testing cohort and demonstrating robust predictability with an AUC of 0.932 in the external validation cohort. Likewise, Model 7-I achieving an AUC of 0.938, and 0.872 in the testing cohort and external validation cohort respectively. DL features extracted from HRCT scans can effectively predict acute exacerbation phenotype in COPD patients.