The primary aim is to construct individualized treatment rule (ITR) models that predict optimal medication selection for deprescribing to maximize potential benefits and reduce harms to patients. To address this aim, we will collect retrospective EHR data for individuals over age 65 from three participating sites (Duke, Vanderbilt, and Medical University of South Carolina) and link this with medication dispensing and insurance claims data from the Centers for Medicare and Medicaid Services(CMS). We will apply causal inference methodology and supervised and unsupervised machine learning algorithms to predict both the benefits (reduced falls, cognitive disorders, hospitalizations) and potential
harms (adverse drug withdrawal events) of medication discontinuation. These predictions will be specific to individual patients and various central nervous system (CNS)-acting medication classes. The resulting machine learning models will be integrated into ITR models which will, in turn, support clinical practice by
recommending the medication class most likely to provide benefits, factoring in individual patient characteristics.