Enhancing Risk Stratification for Incident Systolic Heart Failure through Machine Learning and Natural Language Processing.

Background: Clinical guidelines advocate use of validated risk models in patients experiencing heart failure with reduced ejection fraction (HFrEF) to inform prognosis and assist with management. We developed models for worsening HF (WHF) hospitalizations and death within one year of incident HFrEF using data available within electronic health records (EHR).

Methods: Adults with incident HFrEF were identified from 2013 to 2022 within an integrated healthcare delivery system. We developed decision tree-based models to estimate risks of WHF hospitalization and death within one year of the incident HFrEF date. WHF hospitalizations were ascertained using validated natural language processing algorithms. We evaluated the models using cross-validation and measured final performance (i.e., model discrimination using area under the curve [AUC] and model calibration using the Brier score and calibration plots) on a contemporary hold-out test set of patients from 2021-2022.

Results: Among 28,292 adults with incident HFrEF, 17.3% experienced WHF hospitalization and 15.1% all-cause death at one year of follow-up. We observed an AUC of 0.698 (95% CI: 0.682-0.714) for WHF hospitalization and 0.849 (95% CI: 0.836-0.861) for death and calibrated with a wide range of predicted risks. In comparison, a claims-based risk score displayed an AUC of 0.577 (95% CI: 0.570-0.606) for WHF hospitalization and a smaller dynamic range. Of patients classified as high risk for WHF hospitalization, only 12.0% were receiving full guideline-directed medical therapy at 6 months after HFrEF diagnosis.

Conclusion: Risk models derived using EHR-based data elements can predict both 1-year WHF hospitalization and all-cause mortality in adults with incident HFrEF more accurately than claims-based approaches. These models can be used to improve population management and better target personalized strategies of care.