Development and validation of a parsimonious AI-based mortality risk score for heart failure

Development and validation of a parsimonious AI-based mortality risk score for heart failure. Download PDF Abstract Risk stratification in heart failure (HF) supports clinical decisions, yet existing tools face adoption barriers: conventional scores show modest discrimination and depend on specialised tests (i.e., echocardiography), while artificial intelligence (AI) models require rich longitudinal data and infrastructure. Both approaches face challenges in workflow integration, interpretability, and demonstrating clinical utility.

Using 373,389 adults with HF from the UK Clinical Practice Research Datalink (CPRD) Aurum, we developed and validated an AI-based risk model for all-cause mortality and compared it against the MAGGIC score adapted for EHR data (MAGGIC-EHR). Predictive signals from a Transformer model trained on electronic health records (EHRs) were distilled, via SHAP-based feature selection, into a parsimonious 11-variable model, SIMPLE-HF (Simplified Intelligent Mortality Prediction for Longitudinal EHRs in HF patients), using point-of-care variables including age, body mass index, comorbidities, and medications. For 12-month all-cause mortality, SIMPLE-HF achieved superior discrimination (C-index 0.801, 95% CI 0.795–0.806) versus MAGGIC-EHR (0.735, 0.728–0.741) and was well calibrated.

At a 0.60 high-risk threshold, SIMPLE-HF identified 177 true events per 1000 screened versus 89 for MAGGIC-EHR. Performance gains were consistent for secondary outcomes. This approach may help bridge advanced AI and routine practice, though external validation and prospective evaluation are needed before clinical adoption.

Medical Disclaimer: This article is for informational purposes only and does not constitute medical advice. Always consult with qualified healthcare professionals for medical decisions and treatment options.