Deprescribing Decision-Making using Machine Learning Individualized Treatment Rules to Improve CNS Polypharmacy

ABSTRACT Central Nervous System (CNS) polypharmacy is a cumulative exposure to 3 or more medications with CNS-acting properties. CNS medications include anticholinergics, antidepressants, antipsychotics, benzodiazepines, gabapentinoids, hypnotics, opioids, and muscle relaxants. A growing number of older

Americans are prescribed multiple CNS-acting medications and are affected by CNS polypharmacy, which is rapidly becoming a wide-spread geriatric syndrome. CNS polypharmacy leads to adverse outcomes such as physical and cognitive impairment and medication-related falls, serious injuries, hospitalization, diminished quality of life, and even death. Deprescribing, the systematic process of

tapering or stopping medications, offers potential to reduce the burden of CNS medications. However, deprescribing may also result in the exacerbation of underlying symptoms for which CNS medications were originally prescribed. Existing guidelines, such as Beers or STOPP, do not take into consideration

the complex interplay of multiple medications and individual patient’s unique clinical characteristics, which are necessary for individualized, patient-centered care. Essential are individualized deprescribing solutions that account for these complexities. This project will identify varying benefits and harms of deprescribing across individuals and

medications classes. We will then subsequently develop individualized treatment rules (ITRs) for personalized medication discontinuation recommendations based on patients’ unique clinical characteristics to support clinical practice. First, using causal inference methods, we will analyze PCORNet multi-site electronic health record (EHR) data with linked Medicare claims data to quantify the

probability of medication discontinuation based on patients’ observed clinical characteristics. Then we will employ supervised and unsupervised machine learning (ML) methods, such as decision trees and clustering techniques, to accurately predict patient outcomes after medication class discontinuation

based on that patient’s unique clinical characteristics. Our team will leverage prior experience in causal inference and machine learning applied to EHR data to estimate deprescribing's clinical benefits and harms. We'll ultimately integrate these results into a user-friendly clinical interface, ensuring trustworthy ML and transparent model decision-making

processes. As an initial step, we'll engage clinicians, patients, and community partners early in the design phase to assess understanding and trust in ML-based deprescribing guidance. This will prepare us for a future large-scale clinical trial to test the efficacy in achieving deprescribing goals.