AN EXPLAINABLE MACHINE LEARNING ANALYSIS OF GLUCAGON-LIKE PEPTIDE-1 RECEPTOR AGONIST (GLP-1 RA) USE AND OPIOID-RELATED DIAGNOSIS RISK AMONG ADULT MEDICAID BENEFICIARIES
Author(s)
Bo Zhou, PhD1, Bhumiben P. Patel, PhD2, Rolake A. Neba, PharmD3, Munir Gunes Kutlu, PhD4, Anjali Rajadhyaksha, PhD4, Usha Sambamoorthi, MA, PhD5;
1University of North Texas health sciences center, Fort Worth, TX, USA, 2Temple University Lewis Katz School of Medicine, Philadelphia, PA, USA, 3University of North Texas College of Pharmacy, Fort Worth, TX, USA, 4Lewis Katz School of Medicine at Temple University, Philadelphia, PA, USA, 5Temple University School of Pharmacy, Philadea, PA, USA
1University of North Texas health sciences center, Fort Worth, TX, USA, 2Temple University Lewis Katz School of Medicine, Philadelphia, PA, USA, 3University of North Texas College of Pharmacy, Fort Worth, TX, USA, 4Lewis Katz School of Medicine at Temple University, Philadelphia, PA, USA, 5Temple University School of Pharmacy, Philadea, PA, USA
OBJECTIVES: To evaluate whether GLP-1 RA use is associated with opioid-related problem diagnoses among Medicaid beneficiaries with type 2 diabetes and chronic pain using machine learning (ML) methods.
METHODS: A retrospective cohort analysis was conducted using US MarketScan Medicaid claims data (2022-2023). The study sample (N=63,446) comprised adults aged 18-64 years, continuously enrolled with prescription drug coverage, diagnosed with type 2 diabetes, and chronic pain. The primary outcome was any opioid-related problem diagnosis during follow-up, based on ICD-10 codes for opioid dependence, abuse, or opioid use. Baseline predictors included GLP-1 RA use, demographic, social risk, health system factors, comorbidities, and chronic pain during follow-up. An XGBoost classifier was developed using a stratified 70/30 train-test split with inverse probability weighting. Model performance was evaluated using discrimination, classification, and calibration metrics. SHapley Additive exPlanations (SHAP) were used to characterize the relative contribution of GLP-1 RA use and other predictors. Secondary analyses incorporated baseline opioid-related diagnoses.
RESULTS: Overall, 22.2% used GLP-1 RA, and 11.4% experienced an opioid-related diagnosis during follow-up. In ML models excluding baseline opioid history, discrimination was moderate (AUC=0.713), with high negative predictive value (92-93%) and more limited sensitivity (50%). Chronic pain, anxiety, and depression were the strongest contributors to predicted risk. Across models, both with and without baseline opioid history, GLP-1 RA use demonstrated a small protective contribution. When baseline opioid-related diagnoses were included, model discrimination improved substantially (AUC=0.845), with gains in sensitivity (60%) and positive predictive value (59%).
CONCLUSIONS: Among adult Medicaid beneficiaries with substantial clinical and behavioral complexity, opioid-related diagnoses were primarily associated with chronic pain and behavioral health burden. GLP-1 RA use showed a modest, consistently protective effect across model specifications. Explainable ML approaches may support Medicaid policy and care management by improving opioid risk stratification and informing potential broader behavioral health implications of GLP-1 RA therapy.
METHODS: A retrospective cohort analysis was conducted using US MarketScan Medicaid claims data (2022-2023). The study sample (N=63,446) comprised adults aged 18-64 years, continuously enrolled with prescription drug coverage, diagnosed with type 2 diabetes, and chronic pain. The primary outcome was any opioid-related problem diagnosis during follow-up, based on ICD-10 codes for opioid dependence, abuse, or opioid use. Baseline predictors included GLP-1 RA use, demographic, social risk, health system factors, comorbidities, and chronic pain during follow-up. An XGBoost classifier was developed using a stratified 70/30 train-test split with inverse probability weighting. Model performance was evaluated using discrimination, classification, and calibration metrics. SHapley Additive exPlanations (SHAP) were used to characterize the relative contribution of GLP-1 RA use and other predictors. Secondary analyses incorporated baseline opioid-related diagnoses.
RESULTS: Overall, 22.2% used GLP-1 RA, and 11.4% experienced an opioid-related diagnosis during follow-up. In ML models excluding baseline opioid history, discrimination was moderate (AUC=0.713), with high negative predictive value (92-93%) and more limited sensitivity (50%). Chronic pain, anxiety, and depression were the strongest contributors to predicted risk. Across models, both with and without baseline opioid history, GLP-1 RA use demonstrated a small protective contribution. When baseline opioid-related diagnoses were included, model discrimination improved substantially (AUC=0.845), with gains in sensitivity (60%) and positive predictive value (59%).
CONCLUSIONS: Among adult Medicaid beneficiaries with substantial clinical and behavioral complexity, opioid-related diagnoses were primarily associated with chronic pain and behavioral health burden. GLP-1 RA use showed a modest, consistently protective effect across model specifications. Explainable ML approaches may support Medicaid policy and care management by improving opioid risk stratification and informing potential broader behavioral health implications of GLP-1 RA therapy.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
RWD145
Topic
Real World Data & Information Systems
Topic Subcategory
Health & Insurance Records Systems
Disease
SDC: Diabetes/Endocrine/Metabolic Disorders (including obesity)