Machine learning-based prediction of sudden cardiac death in the general population using electronic health record data

Comprehensive Summary

Youssfi et al. developed a machine learning (ML) model to identify individuals at high risk for sudden cardiac death (SCD) within the general population using longitudinal electronic health record (EHR) data. The study utilized a dataset including 17,172,359 drug prescriptions and 1,639,057 hospital diagnoses across 49,700 participants. The model was trained on a derivation cohort of 12,338 SCD cases and 12,338 matched controls from Greater Paris (2011-2015) and validated on both a temporal cohort (11,620 cases; Paris, 2016-2020) and a geographical cohort (892 cases; Seattle, USA, 2013–2021). The CatBoost algorithm was the top-performing model, achieving an area under the curve (AUC) of 0.83 (95% CI: 0.82-0.84) in the derivation cohort and 0.81 (95% CI: 0.80-0.82) in the temporal validation. Notably, the model successfully identified 26% of all SCD cases in Paris and 33% in Seattle within its highest-risk decile, even though 25.7% of subjects had zero recorded cardiovascular diagnoses in the five years preceding their death. The model demonstrated high specificity for SCD and was notably not predictive of myocardial infarction (MI).

Outcomes and Implications

AI-driven analysis of EHR data can detect subtle "small signals" across a patient's medical history to identify SCD risk that traditional clinical assessments may miss, particularly in the 25% of cases lacking prior cardiovascular symptoms. While the model showed strong temporal stability, its reduced performance in the geographical validation (AUC: 0.66) suggests that population differences or data-encoding variations require local model retraining for optimal usage. Future adoption could allow for targeted preventive strategies, such as earlier cardiology referrals or ICD placement, in the general population. However, prospective trials are necessary to confirm clinical impact on survival rates.