Enhancing risk stratification for incident systolic heart failure through machine learning and natural language...

Area of Science:

• Cardiology
• Medical Informatics
• Predictive Analytics

Background:

• Clinical guidelines recommend validated risk models for heart failure with reduced ejection fraction (HFrEF) to guide prognosis and management.
• Electronic health records (EHR) offer a rich data source for developing such models.
• Previous models often relied on less comprehensive claims data.

Purpose of the Study:

• To develop and validate risk prediction models for worsening heart failure (WHF) hospitalizations and all-cause mortality within one year of incident HFrEF.
• To compare the performance of EHR-based models against traditional claims-based risk scores.

Main Methods:

• Adults with incident HFrEF were identified from 2013-2022 in an integrated healthcare system.
• Decision tree-based models were developed using EHR data to predict 1-year risk of WHF hospitalization and death.
• Model performance was evaluated using cross-validation and a hold-out test set (2021-2022), assessing discrimination (AUC) and calibration (Brier score).
• WHF hospitalizations were identified using validated natural language processing (NLP) algorithms.

Main Results:

• Among 28,292 HFrEF patients, 17.3% experienced WHF hospitalization and 15.1% died within 1 year.
• EHR-based models achieved an AUC of 0.698 for WHF hospitalization and 0.849 for death.
• Claims-based scores showed lower discrimination (AUC 0.577 for WHF hospitalization).
• A significant proportion of high-risk patients (12.0%) were not on guideline-directed medical therapy at 6 months post-diagnosis.

Conclusions:

• EHR-derived risk models offer superior accuracy in predicting 1-year WHF hospitalization and mortality in HFrEF patients compared to claims-based methods.
• These validated models can enhance population health management strategies.
• The models facilitate targeted personalized care interventions for HFrEF patients.