부산대학교 도서관

Introduction: Pulmonary embolism (PE) is a significant cause of mortality in the United States. We aimed to develop deep learning models using computed tomographic pulmonary angiography (CTPA) and clinical data to predict mortality in patients with PE.Hypothesis: Deep learning models based on CTPA and clinical data will outperform PE Severity Index (PESI) in predicting PE mortality.Methods: A total of 927 patients (163 deceased, median age 64, range 13-99, 52% female) and 3978 total CTPAs were identified via retrospective review across three institutions. Data from one institution were randomly split 7:1:2 into training, validation, and internal testing sets. Data from the two remaining institutions were used as an external testing set. Imaging features extracted from PEnet and clinical data (PESI variables) were used to train Random Survival Forest models. Performance was evaluated with concordance index (c-index) and compared to PESI predictions with the Wilcoxon signed-rank test. Kaplan-Meier analysis was performed by stratifying patients into high- and low-risk groups based on the combined imaging and clinical model prediction.Results: Models based on (a) imaging, (b) clinical, and (c) combined imaging and clinical data achieved c-index values of 0.761, 0.774, and 0.843, respectively, on the internal testing set, and 0.660, 0.713, and 0.722 on the external testing set. For both data sets, the combined model outperformed PESI (0.740 and 0.709, respectively, both p<0.001). When stratifying patients into high- and low-risk groups, mortality outcomes were significantly different (p<0.001, Figure).Conclusions: Deep learning models based on combined CTPA features and clinical data outperform PESI for mortality prediction in PE. The addition of imaging to clinical features improves performance compared to clinical features alone.