Abstract Rationale Acute pulmonary embolism (PE) is the third leading cause of cardiovascular death, with 30-day mortality rates of 5-15%, increasing to 28-30% in high-risk patients. The Pulmonary Embolism Severity Index (PESI) is highly sensitive (91%) in identifying low-risk individuals but has limited specificity (41%) for high-risk cases. Although AI and deep learning have been investigated for PE diagnosis, their prognostic potential remains underused. This study aims to apply deep learning to computed tomography pulmonary angiography (CTPA) to predict 30-day mortality in patients with acute PE. Methods This retrospective, multicenter study involved patients diagnosed with acute PE at two academic medical centers. The primary outcome was all-cause mortality within 30 days of PE diagnosis. Pulmonary trunk diameter and right ventricle-to-left ventricle (RV/LV) ratio were measured manually from CTPA images. A deep learning model was developed that combined imaging features extracted from CTPA with radiological parameters to predict patient outcomes. Model performance was assessed using accuracy, precision, and F1-score, and was compared to PESI scores. Results A total of 282 patients were included from two centers (mean age 62.4 years, 54.6% male, 30-day mortality rate 16.0%). The deep learning model using imaging alone achieved an accuracy of 73% (with a precision of 73% and F1-score of 71%). Incorporating radiological measurements improved the accuracy to 80% (precision 80%, F1-score 80%; p = 0.012 vs. imaging alone). In comparison, PESI achieved an accuracy of 54% (precision 26%, F1-score 41%) for outcome prediction in our cohort, significantly underperforming the deep learning model (p = 0.002). Conclusion Our deep learning model, which uses CTPA imaging and radiological measurements, significantly outperforms PESI in predicting 30-day mortality for acute PE. We have demonstrated that combining deep learning algorithms with real-world clinical data improves prognostic accuracy and could support improved patient management strategies in treating acute PE. This abstract is funded by: None
Aktas et al. (Fri,) studied this question.