Generalizable prediction of medication adherence in heart failure

Heart failure (HF) is associated with high rates of hospitalization and mortality. While a number of evidence-based therapies have been shown to improve outcomes for patients with HF, nearly half of these patients are not regularly taking their medications. Although medication adherence can be improved through timely interventions, it is challenging for clinicians to accurately identify and predict medication non-adherence at the point of care. The challenge persists partly because medication adherence is a complex process influenced by an interplay of a multitude of patient-, provider-, system-, community-, and therapy-related factors.

This gap in identifying patients at risk of non-adherence can be addressed through increasing availability of relevant data from electronic health records (EHRs), which affords the potential to make accurate, real time predictions of adherence in HF. In particular, recent linkages of EHR and pharmacy data has created opportunity for incorporation of prior medication fills into EHR-based adherence prediction models that are updated continuously. Using machine learning (ML) techniques with such data allows for incorporation of a large number of intercorrelated risk factors and their interactions into models and for accommodating continuous updates as new information becomes available. Our objective is to build a ML-based algorithm to predict adherence among patients with HF.

The specific aims are: 1) to develop supervised ML algorithms to predict medication adherence among HF patients, using EHR clinical data, linked pharmacy fill data, and location-based data from a large, urban health system; 2) to assess fairness of the developed algorithms by evaluating cross-validated prediction and calibration on key patient subgroups; and 3) to assess generalizability of the algorithms through validation in a second large, urban health system. Our approach is innovative and novel in several ways. First, we will take advantage of linkages between pharmacy fill information and the EHR to incorporate pharmacy data in our models. Second, we utilize geocoding of patient addresses combined with publicly available data to incorporate neighborhood-level variables, which are among the most important predictors of adherence, into our models.

Third, we will ensure generalizability of the prediction algorithm by developing it in one health system and validating the algorithm in a second health system. These models will be developed such that they can be used for point-of-care adherence prediction. Our long term goal is to be able to implement them into the EHR, at which point they can be incorporated into interventions to address medication adherence and, ultimately, improve both adherence and clinical outcomes for patients with HF.

Grant Number: 5R01HL155149-05
NIH Institute/Center: NIH
Principal Investigator: Saul Blecker