COMPASS: A GUIDELINE-CONSTRAINED MULTI-POLICYSUPPORT SYSTEM FOR MEDICATION OPTIMIZATION IN TYPE 2 DIABETES WITH MULTI-MORBIDITY
Author(s)
Xingwei Wu, PhD1, Lijun Wang, PhD2, Zhenglin Yang, PhD3.
1Sichuan Provincial People's Hospital, Chengdu, China, 2Guangxi Medical University, Nanning, China, 3Sichuan Academy of Medical Sciences & Sichuan Provincial People’s Hospital, Chengdu, China.
1Sichuan Provincial People's Hospital, Chengdu, China, 2Guangxi Medical University, Nanning, China, 3Sichuan Academy of Medical Sciences & Sichuan Provincial People’s Hospital, Chengdu, China.
OBJECTIVES: Medication management in patients with type 2 diabetes (T2D) and multimorbidity requires coordinated decisions across therapeutic domains and must balance multiple treatment goals, making decision-making difficult. In this study, we aimed to develop and validate COMPASS, a medication decision-support system based on a multi-policy model, for individualized prescribing in patients with T2D and comorbidities.
METHODS: COMPASS focuses on medication decision support for five coordinated treatment tasks: glucose-lowering therapy, lipid-lowering therapy, blood-pressure control, anticoagulation, and antiplatelet therapy. The system integrates imitation learning to initialize task-specific prescribing policies, Short-RL to optimize 3-month treatment response, and Long-RL to optimize predicted 5-year health value. An evidence-constrained rule engine was also incorporated to exclude contraindicated drugs and combinations. COMPASS was developed and evaluated using 92,552 hospitalization records from 50,017 patients across 25 hospitals in China. External validation was performed in UK Biobank, and COMPASS was further compared head-to-head with three general large language models.
RESULTS: After imitation learning, Short-RL, Long-RL, and evidence-based constraints, COMPASS achieved lower predicted 5-year risk than clinician prescribing in both the external test set of the Chinese cohort and UK Biobank cohort (0.345 [95% CI: 0.339-0.351] vs 0.130 [0.127-0.133] in the external test set; 0.0290 [0.0286-0.0294] vs 0.0113 [0.0113-0.0117] in the UK Biobank cohort). In the head-to-head comparison, COMPASS had a lower predicted 5-year weighted risk than the three large language models (0.218 [0.200-0.236] for COMPASS vs 0.273 [0.256-0.290] for DeepSeek-V4-Pro, 0.278 [0.261-0.294] for Claude Opus 4.7, and 0.277 [0.261-0.294] for ChatGPT 5.5).
CONCLUSIONS: COMPASS provides a structured framework for medication optimization in patients with T2D and multimorbidity, integrating patient-level risk estimation, long-term health-value optimization, and explicit safety constraints. Prospective studies are needed to determine its clinical effectiveness and implementation value in real-world care.
METHODS: COMPASS focuses on medication decision support for five coordinated treatment tasks: glucose-lowering therapy, lipid-lowering therapy, blood-pressure control, anticoagulation, and antiplatelet therapy. The system integrates imitation learning to initialize task-specific prescribing policies, Short-RL to optimize 3-month treatment response, and Long-RL to optimize predicted 5-year health value. An evidence-constrained rule engine was also incorporated to exclude contraindicated drugs and combinations. COMPASS was developed and evaluated using 92,552 hospitalization records from 50,017 patients across 25 hospitals in China. External validation was performed in UK Biobank, and COMPASS was further compared head-to-head with three general large language models.
RESULTS: After imitation learning, Short-RL, Long-RL, and evidence-based constraints, COMPASS achieved lower predicted 5-year risk than clinician prescribing in both the external test set of the Chinese cohort and UK Biobank cohort (0.345 [95% CI: 0.339-0.351] vs 0.130 [0.127-0.133] in the external test set; 0.0290 [0.0286-0.0294] vs 0.0113 [0.0113-0.0117] in the UK Biobank cohort). In the head-to-head comparison, COMPASS had a lower predicted 5-year weighted risk than the three large language models (0.218 [0.200-0.236] for COMPASS vs 0.273 [0.256-0.290] for DeepSeek-V4-Pro, 0.278 [0.261-0.294] for Claude Opus 4.7, and 0.277 [0.261-0.294] for ChatGPT 5.5).
CONCLUSIONS: COMPASS provides a structured framework for medication optimization in patients with T2D and multimorbidity, integrating patient-level risk estimation, long-term health-value optimization, and explicit safety constraints. Prospective studies are needed to determine its clinical effectiveness and implementation value in real-world care.
Conference/Value in Health Info
2026-09, ISPOR Asia Pacific 2026, Bangkok, Thailand
Value in Health, Volume 55, Issue S1
Code
PCR2
Topic
Patient-Centered Research
Disease
SDC: Cardiovascular Disorders (including MI, Stroke, Circulatory), SDC: Diabetes/Endocrine/Metabolic Disorders (including obesity), STA: Personalized & Precision Medicine