Authors
Michael Van Ness, Alexander T Sandhu, Tomas M Bosschieter, Natasha Din, Andrew Ambrosy, Minh Nguyen, Madeleine Udell
Published in
Journal of cardiac failure. Volume 32. Issue 11. Pages 1943-1953.
Abstract
Heart failure (HF) is largely preventable. Contemporary therapies can reduce incident HF among at-risk individuals, making scalable and accurate identification of high-risk patients a priority. Risk prediction models are traditionally developed from long-term epidemiologic cohorts, but machine learning (ML) can derive interpretable models directly from electronic health record (EHR) data captured in routine care. We developed an interpretable ML model for incident HF from health-system EHR data and compared it with established risk equations.
We developed ML-based models for 5-year incident HF using EHR data from the Stanford Healthcare system, incorporating over 2000 predictors including structured data, temporal transformations, and variable interactions. We applied a neural additive model that combines feature selection and nonparametric survival analysis to develop the AIHF-Base model. We removed implausible predictors in an adjusted model (AIHF-Refined) to enhance clinical interpretability. We internally validated the models using a holdout Stanford sample and externally validated them in an independent national Veterans Affairs (VA) cohort with 10-year follow-up. Model discrimination was compared between the AIHF models and two benchmark HF prediction models, PCP-HF and PREVENT-HF.
The AIHF-Base model achieved an AUC of 0.847 in the Stanford cohort and 0.750 in the VA cohort. The AIHF-Refined model had an AUC of 0.840 (Stanford) and 0.741 (VA). Discrimination was comparable to that of PREVENT-HF and PCP-HF in both cohorts; in the VA cohort the AUC was 0.764 for PREVENT-HF and 0.738 for PCP-HF. Of the 14 features in the AIHF-Refined model, representing 13 distinct clinical variables, 7 used longitudinal EHR data. Age and smoking were the only features used by all models. Of 27 total features across models, 8 were used in at least one AIHF and one benchmark model, 19 were used only by AIHF or benchmark models.
Interpretable ML models built from routine EHR data identified clinically coherent predictors of incident HF and achieved discrimination comparable to established risk equations. By drawing on data already captured in clinical care, including longitudinal trends, this approach offers a scalable, transparent strategy for identifying patients at risk for HF and targeting them for prevention.
PMID:
43009248
Bibliographic data and abstract were imported from PubMed on 11 Nov 2026.
Read full publication at:
Please sign in
to see all details.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 6
- Comments 0