(SLA-068) Potential High-risk Medication Clusters Prior to Unplanned Hospitalization Among Older Adults: A Latent Dirichlet Allocation Approach Using Nationwide Claims Data

Background: Unplanned hospitalizations (UH) are common among older adults, significantly affecting clinical outcomes and healthcare utilization. While prior studies have linked polypharmacy and potentially inappropriate medications (PIMs) to increased hospitalization risk, conventional epidemiological methods inadequately capture the complexity of older adults' medication profiles, necessitating advanced analytical approaches such as machine learning techniques.

Objectives: To identify potential high-risk medication clusters, with particular focus on symptom relief and psychotropic medications, among older adults prior to UH using Latent Dirichlet Allocation (LDA).

Methods: Using Taiwan's National Health Insurance Research Database, we identified incident UH events in 2019, then selected patients aged ≥65 years. A novel cluster modeling method was developed comprising four steps: (1) preprocessing prescription records for symptom relief and psychotropic medications based on Anatomical Therapeutic Chemical (ATC) 5th-level codes; (2) identifying latent medication clusters using LDA; (3) optimizing model parameters based on topic coherence scores and clinical interpretability; and (4) visualizing and labeling medication clusters through word clouds, qualitative assessment, and linkage with patient characteristics to delineate the medication clusters.

Results: Using LDA, nine distinct medication clusters were identified among 60,000 older adults (mean age 78.8 ± 8.3 years; 51.5% female) with 407 unique ATC codes. The clusters were characterized by predominant medication types: gastrointestinal, dizziness, musculoskeletal pain, acetaminophen, respiratory/acute, respiratory/mixed, chronic pain, antihistamine, and psychotropics, many of which align with common PIM categories. Significant associations with patient characteristics were also observed in each cluster. For example, the dizziness-related cluster had a higher prevalence of stroke (23.1%) and transient ischemic attack (5.8%); the chronic pain cluster was associated with increased rates of osteoporosis (5.4%), osteoarthritis (22.1%), and fractures (19.8%); and the psychotropics cluster showed elevated rates of anxiety (11.2%), depression (5.0%), and insomnia (14.7%).

Conclusions: This novel, data-driven LDA approach demonstrates significant potential for capturing the intricate heterogeneity of medication utilization patterns in older adults while maintaining clinical interpretability, thereby offering new insights for managing the complex challenges of polypharmacy in geriatric care.