Machine learning based prediction of medication adherence in heart failure using large electronic health record cohort with linkages to pharmacy-fill and neighborhood-level data

Comprehensive Summary

In this study by Adhikari et. al., machine learning (ML) models were created and internally validated for their ability to predict the consistency of adherence to guideline-directed medication therapies (GDMTs) by patients with heart failure (HF). Both electronic health record (EHR) predictors, like pharmacy fill information, and linked neighborhood-level data predictors were used. A cohort of 34,697 patients with HF, at least 1 clinical encounter, and an active prescription for chronic use of GDMTs (greater than 28 days) was employed. The predictions of the average proportion of days covered (PDC) for GDMTs from several ML models were combined using an ensemble model called superlearner. As a result, the mean absolute error (MAE) of superlearner and quantile random forest were the lowest (18.9% for both). The findings suggest potential for employing ML models with EHR-pharmacy data to stratify patient adherence to HF GDMTs. The most significant predictors were PDC prior to a major HF event, and patient characteristics during the event, including frequency of labs and vitals.

Outcomes and Implications

While GDMTs have improved outcomes of patients with HF, a substantial number of patients do not refill their medications. Through the proposed ML model, clinicians would be able to better recognize risk patterns for nonadherence and potentially intervene. Prior to the clinical implementation of the proposed model, stratification of patients across various PDC levels, fairness assessments, and external validations must be conducted. Additionally, factors like patients’ self-efficacy, motivation, and expectations should be considered due to their effects on adherence.