Abstract Body (Do not enter title and authors here): Background and Aims: Obstructive coronary artery disease (CAD) can lead to myocardial infarction or cardiac death. The accuracy of conventional risk prediction models is limited, leading to excessive or insufficient prediction probabilities for patients and resulting in unnecessary angiography. We developed a deep learning (DL) model using non-invasive fundus photographs (FP) to identify obstructive CAD.
Methods and results: In this multicenter cohort study, six different multi-task DL models were developed and validated using patient FPs to detect obstructive CAD, and subsequently compared with logistic regression and guideline-recommended traditional models.. Area under the receiver operating characteristic curve (AUC) was used to evaluate model performance in internal test and independent external test data. The best performing DL model, Inception-Resnet-V2, achieves AUC of 0.838 (95% [CI] 0.799 - 0.870) in the internal test group and 0.769 (95% [CI] 0.737 - 0.793) in the external test group. The net reclassification index showed Inception-Resnet-V2 model predicted better accuracy than traditional models (updated Diamond-Forrester method=0.313, Duke clinical score = 0.397, logistic regression = 0.220; all P < 0.001). Interpretability experiments suggest that the model may be related to retinal vascular characteristics and that the diagnostic efficacy is not affected by traditional risk factors.
Conclusion: The study emphasizes that the non-invasive FP-based DL model outperforms traditional models in predicting obstructive CAD, enabling clinicians to optimize treatment options for patients.