A GUIDELINE-ALIGNED DEEP REINFORCEMENT LEARNING SYSTEM FOR SHORT- AND LONG-TERM THERAPEUTIC OUTCOME OPTIMIZATION IN ANTIDIABETIC THERAPY...
Author(s)
Lijun Wang1, Xingwei Wu, Dr2.
1Nanning, China, 2Sichuan Provincial People’s Hospital, University of Electronic Science and Technology of China, Chengdu, China.
1Nanning, China, 2Sichuan Provincial People’s Hospital, University of Electronic Science and Technology of China, Chengdu, China.
OBJECTIVES: To develop and evaluate an Antidiabetic Decision Support System (ADSS) for individualized medication management in type 2 diabetes (T2D), integrating real-world prescribing patterns, short- and long-term effect optimization, and guideline-based safety constraints.
METHODS: ADSS was developed through imitation learning, deep reinforcement learning, and guideline-constrained recommendation. Imitation learning (behaviour cloning) was used to reproduce physician prescribing across real-world antidiabetic regimens; deep reinforcement learning optimized treatment policies according to simulated long-term health gains; and guideline constraints enforced recommendation and contraindication rules. The system was trained using 22,211 visit records from 10,440 patients across 12 centers and externally validated using 11,096 visit records from 4,848 patients across six centers. Clinical utility was assessed in an observational study of 42 hospitalized patients with T2D at a tertiary hospital in China.
RESULTS: In external validation, ADSS achieved a weighted mean AUC of 0.88 after learning 169 real-world antidiabetic regimens. AUC exceeded 0.80 for 71.3% of regimens, covering 84.2% of visits. Compared with physician-prescribed regimens, reinforcement learning-optimized regimens increased simulated 20-year cumulative health gains by 23.6% (8.48 ± 0.39 vs. 6.54 ± 0.97; P < 0.001). In the clinical utility study, 19 of 21 ADSS recommendations were adopted. Compared with usual care, ADSS-guided treatment was associated with lower 24-hour mean glucose and higher time in range (P < 0.01), while the increase in simulated 20-year cumulative gains was not significant (P = 0.14). Hypoglycaemia occurred in 4 of 19 ADSS-guided regimens and 6 of 21 control regimens; one severe hypoglycaemic event occurred in the control group.
CONCLUSIONS: ADSS may support individualized and guideline-consistent treatment decisions for patients with T2D. It showed robust validation performance, improved simulated long-term health gains, and favourable short-term glycaemic outcomes. Larger prospective studies are needed to confirm its clinical effectiveness and safety.
METHODS: ADSS was developed through imitation learning, deep reinforcement learning, and guideline-constrained recommendation. Imitation learning (behaviour cloning) was used to reproduce physician prescribing across real-world antidiabetic regimens; deep reinforcement learning optimized treatment policies according to simulated long-term health gains; and guideline constraints enforced recommendation and contraindication rules. The system was trained using 22,211 visit records from 10,440 patients across 12 centers and externally validated using 11,096 visit records from 4,848 patients across six centers. Clinical utility was assessed in an observational study of 42 hospitalized patients with T2D at a tertiary hospital in China.
RESULTS: In external validation, ADSS achieved a weighted mean AUC of 0.88 after learning 169 real-world antidiabetic regimens. AUC exceeded 0.80 for 71.3% of regimens, covering 84.2% of visits. Compared with physician-prescribed regimens, reinforcement learning-optimized regimens increased simulated 20-year cumulative health gains by 23.6% (8.48 ± 0.39 vs. 6.54 ± 0.97; P < 0.001). In the clinical utility study, 19 of 21 ADSS recommendations were adopted. Compared with usual care, ADSS-guided treatment was associated with lower 24-hour mean glucose and higher time in range (P < 0.01), while the increase in simulated 20-year cumulative gains was not significant (P = 0.14). Hypoglycaemia occurred in 4 of 19 ADSS-guided regimens and 6 of 21 control regimens; one severe hypoglycaemic event occurred in the control group.
CONCLUSIONS: ADSS may support individualized and guideline-consistent treatment decisions for patients with T2D. It showed robust validation performance, improved simulated long-term health gains, and favourable short-term glycaemic outcomes. Larger prospective studies are needed to confirm its clinical effectiveness and safety.
Conference/Value in Health Info
2026-09, ISPOR Asia Pacific 2026, Bangkok, Thailand
Value in Health, Volume 55, Issue S1
Code
HSD16
Topic
Health Service Delivery & Process of Care
Disease
No Additional Disease & Conditions/Specialized Treatment Areas, SDC: Diabetes/Endocrine/Metabolic Disorders (including obesity)