Explainable Multimodal Early-Warning Analytics for Preventing Falls and Avoidable Hospital Transfers in Long-Term Care: A FHIR-Native and Fairness-Aware Framework

Falls, emergency department transfers, unplanned hospital admissions and 30-day readmissions are recurrent threats to safety, continuity and quality in long-term care. Existing risk tools commonly depend on single-source assessments, static scores or hospital-centric data models that do not exploit the temporal and multimodal information generated in nursing homes. This study develops and evaluates an explainable, FHIR-native and fairness-aware early-warning framework that integrates demographics, diagnoses, vital signs, medication burden, activities of daily living, prior utilization, mobility indicators and natural-language-derived change-of-condition signals. Because no single openly accessible Kaggle dataset contains all long-term-care modalities and all four outcomes, the empirical demonstration uses a reproducible 12,000-resident synthetic cohort calibrated to distributions and relationships reported in the Kaggle Diabetes 130-US Hospitals readmission dataset, an elderly fall-prediction dataset and peer-reviewed long-term-care literature. Facility-level holdout validation compared logistic regression, random forest and gradient boosting models. Performance was assessed with AUROC, area under the precision-recall curve, Brier score, sensitivity, specificity and calibration. Fairness was evaluated across sex, race and dual-eligibility groups at a capacity-constrained top-quintile alert threshold. The best-performing models achieved useful discrimination across outcomes, while calibration and subgroup analysis exposed operational trade-offs that would be hidden by accuracy alone. The framework maps inputs and outputs to FHIR resources and specifies an alert explanation packet comprising predicted risk, dominant drivers, uncertainty, trend context and recommended human review. The results support a governance model in which predictive analytics supplements, rather than replaces, nursing assessment and is continuously monitored for drift, workload effects and inequitable error profiles. The proposed architecture offers a practical foundation for prospective validation in skilled nursing facilities and for interoperable deployment across electronic health-record vendors.