[Development of a deep learning model for predicting adverse cardiovascular events in patients with acute coronary syndrome based on retinal fundus images].

Objective: To develop and validate a deep learning model based on retinal fundus images for predicting the risk of long-term adverse cardiovascular events in patients with acute coronary syndrome (ACS). Methods: This multicenter diagnostic study enrolled ACS patients from July 2021 to July 2023 at Beijing Anzhen Hospital, Beijing Daxing Hospital, Beijing Miyun Hospital, and the Second Affiliated Hospital of Hainan Medical University. Retinal fundus images were acquired. Patients underwent follow-up, with the primary endpoint defined as major adverse cardiovascular and cerebrovascular events (MACCE). Patients were divided into a development set and an external test set. A multimodal deep learning model based on Inception-Resnet-V2 was constructed. Predictive performance for MACCE was evaluated using the area under the receiver operating characteristic curve (AUC). Patients were further stratified into low-and high-risk groups according to the optimal cutoff value derived from the receiver operating characteristic curve, and Kaplan-Meier survival curves were generated. Class activation mapping was used to analyze the retinal regions the model focused on during decision-making. Results: A total of 4 703 ACS patients were included, with a median age of 59 years and 3 504 (74.51%) males. The development set comprised 1 521 patients, and the external validation set included 3 182 patients. The maximum follow-up duration was 883 days. The MACCE incidence rate was 27.42% (417/1 521) in the development cohort and 10.03% (319/3 182) in the external validation cohort. The model's AUC for predicting MACCE was 0.709 (95%CI 0.676-0.742) in the development cohort and 0.644 (95%CI 0.613-0.675) in the external validation cohort. Kaplan-Meier survival curves demonstrated that the MACCE incidence was higher in the high-risk group than in the low-risk group in both the development and external validation cohorts (P<0.05 for log-rank tests). Class activation mapping analysis revealed that the model focused on retinal vascular regions during decision-making. Conclusions: A deep learning model based on retinal fundus images effectively predicts long-term adverse cardiovascular event risk in ACS patients, demonstrating robust risk stratification capability and clinical interpretability. It holds promise as a noninvasive risk assessment tool.