Development and validation of a neural network-based survival model for mortality in ischemic heart disease

Background Current risk prediction models for ischemic heart disease (IHD) use a limited set of established risk factors and are based on classical statistical techniques. Using machine-learning techniques and including a broader panel of features from electronic health records (EHRs) may improve prognostication. Objectives Developing and externally validating a neural network-based time-to-event model (PMHnet) for prediction of all-cause mortality in IHD. Methods We included 39,746 patients (training: 34,746, test: 5,000) with IHD from the Eastern Danish Heart Registry, who underwent coronary angiography (CAG) between 2006-2016. Clinical and genetic features were extracted from national registries, EHRs, and biobanks. The feature-selection process identified 584 features, including prior diagnosis and procedure codes, laboratory test results, and clinical measurements. Model performance was evaluated using time-dependent AUC (tdAUC) and the Brier score. PMHnet was benchmarked against GRACE Risk Score 2.0 (GRACE2.0), and externally validated using data from Iceland (n=8,287). Feature importance and model explainability were assessed using SHAP analysis. Findings On the test set, the tdAUC was 0.88 (95% CI 0.86-0.90, case count, cc=196) at six months, 0.88(0.86-0.90, cc=261) at one year, 0.84(0.82-0.86, cc=395) at three years, and 0.82(0.80-0.84, cc=763) at five years. On the same data, GRACE2.0 had a lower performance: 0.77 (0.73-0.80) at six months, 0.77(0.74-0.80) at one year, and 0.73(0.70-0.75) at three years. PMHnet showed similar performance in the Icelandic data. Conclusion PMHnet significantly improved survival prediction in patients with IHD compared to GRACE2.0. Our findings support the use of deep phenotypic data as precision medicine tools in modern healthcare systems.