Development and Validation of Scalable Predictive Analytics Frameworks for Healthcare Information Systems and Enterprise Decision Support

Abstract

Healthcare organizations increasingly require predictive analytics that can operate across heterogeneous information systems, preserve statistical validity under temporal and institutional change, and translate risk estimates into governable enterprise decisions. This quantitative simulation study developed and validated a scalable predictive analytics framework for healthcare information systems and enterprise decision support using a synthetic multi-site electronic-health-record benchmark. The benchmark contained 130,000 discharge encounters distributed across 13 simulated U.S. health systems, four geographic regions, and five annual periods from 2021 through 2025. No real patient data were used. Eighteen predictors represented demographics, coverage, prior utilization, comorbidity, index-encounter acuity, laboratory abnormality, length of stay, medication burden, discharge disposition, social risk, digital engagement, care coordination, service line, and enterprise context. The reference decision-support outcome was 30-day unplanned acute-care reutilization, with an overall simulated event rate of 18.0%. Development used 50,305 observations from nine health systems during 2021–2023, followed by same-system validation, temporal holdout, geographic holdout, and an independent 2025 test across four unseen systems. Logistic regression, Decision Tree, Naive Bayes, Random Forest, Gradient Boosting, neural-network, and XGBoost models were compared with discrimination, calibration, class-specific, robustness, and subgroup metrics. XGBoost achieved the strongest independent-test ROC-AUC of 0.816 (bootstrap 95% CI, 0.804–0.825), PR-AUC of 0.591, accuracy of 79.5%, sensitivity of 64.3%, specificity of 83.1%, F1-score of 54.8%, and Brier score of 0.115 at the validation-selected operating point. At a lower operational threshold of 0.15, sensitivity increased to 72.5% with specificity of 75.6%, demonstrating the importance of workflow-specific threshold selection. The model retained ROC-AUC values of 0.836 in temporal holdout and 0.833 in geographic holdout, while independent-test calibration slope was 0.93. Explainability analysis ranked acuity, prior admissions, comorbidity, laboratory abnormality, care coordination, and discharge disposition among the dominant features. The framework further integrated FHIR/OMOP-oriented data standardization, model registry and version control, subgroup monitoring, drift surveillance, human review, and auditable decision-support interfaces. Results show that scalable healthcare prediction is not primarily a model-selection problem: enterprise value depends on interoperable data engineering, external validation, calibration, threshold governance, explainability, monitoring, and explicit human accountability. Because the benchmark is synthetic, the findings validate the analytical architecture and implementation logic rather than clinical effectiveness and require prospective real-world validation before patient-care use.