Synergistic Agentic AI and Data-Driven Optimization for Polypharmacy Management
Description
Study description and research hypothesis. This study investigates a synergistic approach that integrates structured polypharmacy analysis and optimization, supported by PM-TOM (Personalized Medicine–Treatment Optimization Method), with a multi-agent system (MAS) of specialist LLM agents. The research hypothesis is that combining agentic AI clinical reasoning with patient-specific analysis and optimization can identify alternative polypharmacy treatments with fewer adverse drug reactions (ADRs), drug interactions, and potentially inappropriate medication risks, while preserving human clinical judgment. Data sources and case study. Data sources include the published deprescribing case of Farrell, Shamji, and Dalton involving a 79-year-old man referred to a geriatric day hospital for cognition, pain, falls, impaired mobility, mood, dizziness, and insomnia; AI-generated treatment recommendations and drug alternatives; and PM-TOM medication safety analyses. The published discharge treatment is the reference (V1). We evaluated three modifications: an AI agent-generated replacement treatment (V2); a PM-TOM-optimized treatment based on AI-generated alternatives (V3); and an AI-assisted, clinically refined treatment based on further review of V3 (V4). Data generation and evaluation. We analyzed each version with PM-TOM (https://www.abs-info-age.com/) using the same patient data and evaluation framework. The dataset contains treatment compositions and patient-specific counts and details of drug–ADR associations, critical and other drug–drug interactions (DDIs), drug–condition interactions (DCIs), drug–food interactions (DFIs), and Beers Criteria risks. Findings. The principal findings show substantial reductions in several clinically relevant risk measures in V2, V3, and V4 compared with V1, particularly critical DDIs, DCIs, DFIs, and Beers Criteria risks. Data interpretation and potential use. Results can be used to examine how specialist LLM agents and structured optimization complement one another, evaluate alternative agent or optimization strategies, and support research on AI-assisted polypharmacy management, deprescribing, and human–AI clinical decision support. Beyond retrospective deprescribing, the combined approach could support prospective medication optimization in primary care, helping prevent PIM prescribing and prescribing cascades before complex deprescribing becomes necessary. Patient-specific, prioritized PM-TOM reports can also help pharmacists and other clinicians monitor and adjust treatment while reducing non-actionable alerts and alert fatigue.