UNIHAA: A Multi-Agent Reinforcement Learning Framework for Autonomous Healthcare Appointment Management – From Behavioral Theory to Agentic Action

Background: No shows are a loss of clinical time, add to waiting lists and break continuity of care. The literature review indicates that the prediction of no-shows is still a significant challenge and reminder and targeted-intervention strategies have been shown to decrease no-shows, but with varying effects depending on the context [25-27]. The new healthcare literature also recognizes agentic and multi-agent AI as the new strategies of coordinated and goal-oriented workflows [1,5]. The aim of UNIHAA was to transcend the passive risk prediction, and to work towards coordinated, transparent and safety-governed action.Methods: We analysed 3,275,731 eligible non-cancelled outpatient visits of Saudi military healthcare facilities in 2021-2025; three records with missing covariate data were not included in the multivariable regression analysis, resulting in N=3,275,728. The empirical analysis looked at the trends of time, the impact of lead time, the variation of facilities and specialties, and history-conscious forecasting. UNIHAA combines the Individual, Clinical, System and Environmental levels of the Integrated Behavioral Access Network (IBAN) with a 10-agent architecture. It consists of history-sensitive risk stratification, LinUCB contextual-bandit intervention selection, dynamic slot recovery, federated multi-site learning, Page-Hinkley concept-drift detection, PPO-based coordination with Centralized Training Decentralized Execution (CTDE), FHIR interoperability, explainable AI, and human oversight.Results: The overall no-show rate in the retrospective study was 22.6% and the historical rate of no-show has decreased from 30.2% in 2021 to 18.7% in 2025. There was a strong association between extended lead times ( >60 days) and non-attendance (adjusted OR 2.47, 95% CI 2.44-2.51). The administrative-only model had an AUC of 0.605, on the complete data set. The development AUC=0.790 (95% CI 0.789-0.791), temporal holdout AUC=0.782, Brier score=0.171, PR-AUC=0.523 and ECE=0.018 were obtained in the companion AALAM analysis of a separate history-eligible cohort (N=672,441). Due to the different cohorts used for evaluation, the AUCs of the two models are descriptive benchmarks and not a same-cohort effect estimate. An assumption-based LinUCB policy simulation projected a 4.2 percentage-point attendance uplift (28,243 additional attended appointments over two years in the modeled cohort), while dynamic overbooking simulation recovered 12.3% of otherwise lost capacity. Federated prediction achieved AUC=0.785 versus 0.790 under centralized pooled training.Conclusions: UNIHAA extends the previously published AALAM empirical study through an original, theory-informed architecture for AI-driven healthcare appointment decisions. Its proposed contributions include hierarchical multi-agent coordination, contextual-bandit intervention selection, reinforcement-learning-based operational optimization, fairness-constrained decision governance, and human oversight. Retrospective predictive findings and simulation-derived operational projections provide a foundation for system design but do not establish the effectiveness of integrated PPO/CTDE coordination or live autonomous deployment. The proposed fairness target of DPR ≤ 1.05 also requires operational validation. UNIHAA therefore contributes a computational decision-support architecture whose effectiveness, safety, and scalability require prospective evaluation.