Traditional Methods Hold Their Ground Against Machine Learning in Predicting Potentially Inappropriate Medication Use in Older Adults

Abstract

ObjectivesMachine learning methods have gained much attention in health sciences for predicting various health outcomes but are scarcely used in pharmacoepidemiology. The ability to identify predictors of suboptimal medication use is essential for conducting interventions aimed at improving medication outcomes. It remains uncertain whether machine learning methods could enhance the identification of potentially inappropriate medication use among older adults compared with traditional methods. This study aimed to (1) to compare the performances of machine learning models in predicting use of potentially inappropriate medications and (2) to quantify and compare the relative importance of predictors in a population of community-dwelling older adults (>65 years) in the province of Québec, Canada. MethodsWe used the Québec Integrated Chronic Disease Surveillance System and selected a cohort of 1 105 295 older adults of whom 533 719 were potentially inappropriate medication users. Potentially inappropriate medications were defined according to the Beers list. We compared performances between 5 popular machine learning models (gradient boosting machines, logistic regression, naive Bayes, neural networks, and random forests) based on receiver operating characteristic curves and other performance criteria, using a set of sociodemographic and medical predictors. ResultsNo model clearly outperformed the others. All models except neural networks were in agreement regarding the top predictors (sex and anxiety-depressive disorders and schizophrenia) and the bottom predictors (rurality and social and material deprivation indices). ConclusionsIncluding other types of predictors (eg, unstructured data) may be more useful for increasing performance in prediction of potentially inappropriate medication use.