Reinforcement Learning for Chronic Care Pathway Optimization: A Unified Framework across Three Clinical Goal Types
- 1Guangdong Provincial Hospital of Chinese Medicine, Hainan Hospital
Abstract
Chronic care requires sequential treatment under competing biomarker, safety, and cost constraints, yet clinical goal structures differ across diseases. This version asks whether one physiology-informed reinforcement-learning paradigm can be written once and then instantiated for three goal structures-- treat-to-target, multi-marker cruise, and bounded cycle completion--without a disease-specific policy architecture. Gout, chronic kidney disease, and PCOS-mediated fertility treatment are the three instances. Training is behaviour cloning followed by proximal policy optimization on 500 simulated trajectories per disease. Evaluation uses paired seeds (N = 50 primary; N = 500 bootstrap 95% confidence intervals), multi-seed offsets 42-51, ablation, and out-of-distribution profiles.
PCOS (cycle completion): PPO 72.0% against the doctor baseline 54.0% at N = 50; at N = 500, PPO 69.8% [65.6, 73.8] against doctor 52.8% [48.8, 57.2]. Gout (treat-to-target): PPO 88.0% against 90.0% (McNemar p = 1.0). CKD (cruise): doctor 32.0%, BC/PPO 38.0%.
This v2 withdraws three bibliographic coordinates that appeared in v1 and that do not match any recoverable source. The withdrawn comparison is replaced by a NOR-Gout calibration at weeks 4, 8 and 12 that belongs to the recalibrated kernel of the parallel journal manuscript. Shared BC[->]PPO training is instantiated across the three goal types without cross-disease weight sharing. Prospective validation remains necessary.
HighlightsO_LIThree goal types are hosted as different success tests, not as a claimed clinical taxonomy. C_LIO_LIPCOS cycle completion +18 percentage points against the doctor baseline at N = 50. C_LIO_LIGout remains non-inferior to the guideline-mimicking doctor in the v1 evaluation table. C_LIO_LINOR-Gout weeks 4/8/12 belong to the recalibrated kernel, not Section 3. C_LIO_LIBehaviour-cloning warm-start remains necessary under the gout kernel used here. C_LI
Related
- Retail sales optimization using a hybrid K-Means, particle swarm optimization and reinforcement learning framework
- Multiproperty optimization of therapeutic antibodies through sequence editing via large language model and reinforcement learning
- Adaptive intelligent resource optimization in wireless sensor networks using reinforcement learning and compressive sensing
- Continuous control deep reinforcement learning for joint resource optimization in STAR-RIS assisted mmWave-NOMA systems with hardware impairments
- An integrated evolutionary optimization and machine learning framework for predicting pharmacokinetics of nanocarrier-encapsulated drugs
BibTeX
@misc{wang2026reinforcement,
title = {{Reinforcement Learning for Chronic Care Pathway Optimization: A Unified Framework across Three Clinical Goal Types}},
author = {R. Wang and H. Chen and Y. Wu and Z. Li and R. Shen and F. He and S. Zhao and N. Zheng},
howpublished = {medRxiv},
year = {2026},
month = jul,
doi = {10.64898/2026.07.03.26357209},
url = {https://www.medrxiv.org/content/10.64898/2026.07.03.26357209v2},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Reinforcement Learning for Chronic Care Pathway Optimization: A Unified Framework across Three Clinical Goal Types
AU - R. Wang
AU - H. Chen
AU - Y. Wu
AU - Z. Li
AU - R. Shen
AU - F. He
AU - S. Zhao
AU - N. Zheng
PY - 2026
DA - 2026/07/03
PB - medRxiv
DO - 10.64898/2026.07.03.26357209
UR - https://www.medrxiv.org/content/10.64898/2026.07.03.26357209v2
AB - Chronic care requires sequential treatment under competing biomarker, safety, and cost constraints, yet clinical goal structures differ across diseases. This version asks whether one physiology-informed reinforcement-learning paradigm can be written once and then instantiated for three goal structures-- treat-to-target, multi-marker cruise, and bounded cycle completion--without a disease-specific policy architecture. Gout, chronic kidney disease, and PCOS-mediated fertility treatment are the three instances. Training is behaviour cloning followed by proximal policy optimization on 500 simulated trajectories per disease. Evaluation uses paired seeds (N = 50 primary; N = 500 bootstrap 95% confidence intervals), multi-seed offsets 42-51, ablation, and out-of-distribution profiles. PCOS (cycle completion): PPO 72.0% against the doctor baseline 54.0% at N = 50; at N = 500, PPO 69.8% [65.6, 73.8] against doctor 52.8% [48.8, 57.2]. Gout (treat-to-target): PPO 88.0% against 90.0% (McNemar p = 1.0). CKD (cruise): doctor 32.0%, BC/PPO 38.0%. This v2 withdraws three bibliographic coordinates that appeared in v1 and that do not match any recoverable source. The withdrawn comparison is replaced by a NOR-Gout calibration at weeks 4, 8 and 12 that belongs to the recalibrated kernel of the parallel journal manuscript. Shared BC[->]PPO training is instantiated across the three goal types without cross-disease weight sharing. Prospective validation remains necessary. HighlightsO_LIThree goal types are hosted as different success tests, not as a claimed clinical taxonomy. C_LIO_LIPCOS cycle completion +18 percentage points against the doctor baseline at N = 50. C_LIO_LIGout remains non-inferior to the guideline-mimicking doctor in the v1 evaluation table. C_LIO_LINOR-Gout weeks 4/8/12 belong to the recalibrated kernel, not Section 3. C_LIO_LIBehaviour-cloning warm-start remains necessary under the gout kernel used here. C_LI
ER -