ARCHITECTURE FOR HEALTHCARE ANALYTICS

Healthcare organizations increasingly convert analytical outputs into consequential clinical and operational actions: firing deterioration and sepsis alerts, routing triage, prioritizing imaging worklists, allocating beds and staff, and gating prior authorization, often with no clinician in the immediate path. In that setting the discriminatory power of a model, whether measured by AUROC or AUPRC, is no longer a sufficient basis for trust. Failures originate upstream and downstream of the model: in electronic-health-record inputs whose provenance or temporal validity cannot be established, in probabilities that are uncalibrated across sites and subpopulations, in alerts whose rationale cannot be reproduced for the clinician who received them, in policy boundaries drawn in the wrong place, in automated actions that cannot be safely held or reversed, and in feedback loops such as alert fatigue through which the system quietly reshapes the behavior it measures. This paper adopts as its runtime-governance foundation the Full-Stack Production-Platform Reference Architecture introduced by Mesbaul Haque Sazu, in which every layer is contract-bearing and the runtime is wrapped by governance and observability spines [19], and grounds it in the account of big data analytics and artificial intelligence in healthcare management given by Sazu and Jahan [20]. Building directly on both, the paper develops a single decision-to-action architecture for healthcare analytics that keeps a decision-assurance plane and an execution-assurance plane distinct while binding them through a shared evidence contract. It states pre-commit admissibility for a clinical or operational action as a conjunction of individually testable conditions: input provenance and temporal validity, independent validation, policy and residency permissibility, evidence completeness, calibrated confidence gating, reversibility, and remit-based clinician authority. It maps the adopted runtime invariants onto one governed care lifecycle, defines a cumulative conformance model, and specifies a four-level evaluation protocol together with the measurement hazards particular to care: dataset shift across sites and time, label latency with treatment-induced censoring, and alert-fatigue feedback. A synthetic sensitivity analysis of confidence-gated clinician review illustrates the escalation-versus-harm trade-off and the cost surface behind threshold selection; it is illustrative and carries no empirical weight. The contribution offered here is the governance interface between Sazu’s execution-assurance account and the healthcare-analytics context, not a replacement for either.