A Data-Driven Predictive Framework for Reducing Medication Non-Adherence Cost in the U.S. Healthcare System

Abstract

Recent estimates from the healthcare industry peg the annual cost burden due to non-adherence in the United States at USD 290 billion to USD 500 billion [1], [2] of direct costs, which is one of the largest avoidable cost drivers. This paper presents a 5-step data-driven pipeline that takes as input patient demographic, clinical, behavioral and access-related data, and processes it through feature engineering for 18 variables, followed by 3 complementary base learners (logistic regression, random forest and gradient boosting), a soft voting ensemble of the 3 base learners and a final output that includes both a tiered risk classification per patient and an estimated avoidable annual cost under a targeted intervention policy. The framework was tested on a sample of N = 215 adult patients who were on a chronic-disease pharmacotherapy at U.S. ambulatory clinical sites, where 20.0 % had a proportion-of-days-covered (PDC) less than the 0.80 threshold established by the Pharmacy Quality Alliance (PQA) for HEDIS. The soft-voting ensemble had a receiver operating characteristic (ROC) area under the curve (AUC) of 0.839, an accuracy of 0.831, and an F1 score of 0.560, and the five-fold cross validated ROC-AUC was 0.776 on a 30 % stratified hold out set. The five features with the greatest predictive power in both feature importances were health literacy, out-of-pocket copayment, patient self-efficacy, depression burden, and side-effect burden in the gradient boosting and random forest ranking. Under a conservative intervention-effectiveness assumption of 40 % and intervention cost of USD 320 per patient per year, the cost-scenario analysis of the ensemble-flagged non-adherent patients suggests a net annual saving of USD 2,564 per captured patient – a return of the targeted-intervention policy of an order of magnitude. It is interpretable in terms of the ranking of features in Gradient-Boost, relatively easy to implement on top of a standard electronic-health-record pipeline, and it can be audited by its risk tier per patient. The framework will be expanded to a prospective multi-site evaluation and to the fairness-across-subgroups analysis that is becoming a more routine requirement of current U.S. regulatory guidance for future work.