DR-UCSS: A Distributionally Robust, Uncertainty-calibrated Counterfactual Safety-shielded Reinforcement Learning Framework for Hospital Operations

Read full text Summarise with VIRAT
Authors: Leela Prasad Gorrepati, Sadhasivam Mohanadas
Cited by: 0
References: 26

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Hospital administration involves uncertainty due to random arrival of patients, randomness in the severity of patients, and scarcity of clinical resources.Recent developments in statistical modelling, machine learning (ML), and deep learning (DL) have increased the capability of risk adjustment and outcome prediction.Nonetheless, such methods are basically observational and are insufficient for safe, sequential operational control in complex settings such as hospitals.Specifically, risk prediction accuracy alone is not sufficient to ensure the safety of capacity utilization or robustness under sudden demand surges.In practice, rare but severe congestion events can dominate system-level performance and patient outcomes.To address this gap, this work introduces a decision-aware framework that integrates causal modelling with constrained reinforcement learning to enable safe and interpretable hospital operational control.Rather than relying solely on correlational forecasts, the proposed approach explicitly models intervention effects and capacity constraints during sequential decision-making.The utilization of the proposed DR-UCSS framework significantly reduces the waiting time and length of stay of the patients, achieving a reduction of 21.8% in average waiting time and 17.4% in average length of stay compared to the strongest constrained reinforcement learning baseline.The framework also improves system throughput and reduces overload violation probability by 39.6%, alongside a 34.2% reduction in extreme tail-risk (CVaR_0.95) under demand surge conditions.Additional analyses for ablation, sensitivity, and runtime consistently demonstrate that the proposed approach is stable, interpretable, and deployable.Through extensive experimental evaluations with large-scale hospital data, calibrated scenarios, and robustness analyses, DR-UCSS is shown to outperform heuristic scheduling policies, unconstrained reinforcement learning, and classical constrained policy optimization baselines.Beyond forecasting outcomes, the proposed framework actively supports risk-aware operational decisions, enabling hospitals to maintain operational capacity with less than 2.3% safety violations across all tested scenarios while preserving interpretability in real-world deployments.

Fields of Study & Subjects

Healthcare Operations and Scheduling Optimization Emergency and Acute Care Studies Healthcare Quality and Management

Publication details

Journal / source
International journal of intelligent engineering and systems
Volume
19
Issue
4
Pages
286–302
Publication date
23 March 2026
Language
EN