Machine learning approaches to predicting medication nonadherence: a scoping review.

BACKGROUND

Medication nonadherence is a common, preventable cause of adverse clinical outcomes. Predictive models identifying risk of nonadherence could enable proactive intervention.

OBJECTIVE

This scoping review aimed to describe relevant predictors, model training and evaluation processes, and how adherence was classified to inform implementation of clinically actionable models.

MATERIALS AND

METHODS

A systematic search of PubMed, Embase, and Web of Science was conducted for studies published between January 2015 and December 2024 describing creation of models predictive of future medication adherence using machine learning methods. Conference abstracts, review articles, study protocols, or full text articles unavailable to authors were excluded. Data was extracted and study risk of bias assessed by an investigator-specified scale. Quantitative analysis was performed in studies reporting area under receiver operating characteristic curve (AUC), analyzing characteristics of the model with the highest reported AUC ("primary model") per study.

RESULTS

52 studies were included, of which 14 were considered low risk of bias, 34 moderate, and 4 high. 9 did not report AUC and were excluded from quantitative analysis. Adherence was most frequently assessed using indirect, dispense history-based methods such as proportion of days covered. Primary models incorporating diagnostic or subject-reported data had higher median AUC (diagnostic 0.837; subject-reported 0.828; overall 0.82). Common important predictors included the Beliefs about Medicines questionnaire, comorbidities, medication history, prior adherence and socioeconomic factors. Random forest and logistic regression models were identified as the highest performing models most frequently.

CONCLUSION

Approaches to modeling and evaluating adherence prediction were highly variable, however several successful algorithms, predictors, and training techniques were identified. Future research should prioritize operational feasibility and clinical utility in development of predictive models to ensure creation of effective clinical decision support tools.