Integrated Spatio-Temporal Counterfactual Recourse and Closed-Loop Reinforcement Learning for Equitable Endocrine Disease Triage and Diagnostics Distribution in Resource-Constrained Healthcare Networks

• Abilly Elly

  Texas University

  Author

Endocrine disorders present a growing global health burden, yet diagnostic and triage systems in resource-constrained healthcare networks face persistent inequities exacerbated by specialist shortages, infrastructure limitations, and geographically disparate populations. Existing approaches, predominantly relying on static clinical guidelines or isolated machine learning models, fail to address the dynamic interplay of spatial accessibility, temporal disease progression, and counterfactual reasoning required for equitable resource allocation. This research presents a novel integrated framework combining spatio-temporal counterfactual recourse with closed-loop reinforcement learning to optimize endocrine disease triage and diagnostic distribution. The methodology leverages retrospective electronic health records from 15,992 patients across 47 underserved clinics, incorporating geospatial accessibility metrics, temporal disease trajectories, and counterfactual action sequences generated through structural causal models. The framework achieved 89.4% triage accuracy for urgent endocrine conditions, outperforming conventional triage protocols (72.1%, p<0.001) and static machine learning baselines (78.3%). Geospatial resource optimization reduced mean diagnostic wait times from 18.7 to 6.2 days in high-need regions while improving equity indices by 34.2%. The closed-loop reinforcement learning component demonstrated robust performance (94.1% glucose regulation success in simulated diabetes cohorts) without the temporal resampling pitfalls identified in existing offline RL approaches. This framework provides a replicable, ethically-grounded blueprint for equitable endocrine care delivery in underserved healthcare ecosystems.

• pdf

Copyright (c) 2026 Abilly Elly (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.