Search

Skilled nursing facilities (SNFs) in the U.S. serve over 1.5 million residents and experience continuous census volatility driven by admissions, discharges, and mortality, impacting staffing, bed availability, and care quality. Existing forecasting methods rarely capture these dynamics together, leading to reactive and inefficient operational decisions. A need exists for accurate, multi-horizon, and data-integrated forecasting systems. Traditional models like ARIMA and LSTM are limited in SNF census forecasting because they produce single-point estimates, fail to model uncertainty, and cannot effectively integrate heterogeneous data such as facility characteristics, temporal utilization patterns, and external factors like COVID-19 prevalence. They also lack interpretability, reducing their usefulness for decision-making. This study introduces an attention-based Temporal Fusion Transformer (TFT) for multi-horizon SNF census forecasting (1, 7, 14, and 30 days). It integrates admissions, discharges, and COVID-19 prevalence through dedicated encoders and applies variable selection networks, LSTM layers, and multi-head attention to capture temporal dependencies and feature importance. The model outputs quantile forecasts (10th, 50th, 90th percentiles) to quantify uncertainty. The TFT enhances interpretability by identifying which past events and features most influence predictions at each horizon, enabling administrators to understand how admissions trends, discharge patterns, and COVID-19 surges affect census dynamics. The proposed framework enables proactive SNF capacity planning by combining multi-source data with interpretable, uncertainty-aware forecasting, supporting a shift from reactive staffing to anticipatory resource allocation and improved operational efficiency.

During the COVID-19 pandemic, machine learning models developed in high-resource hospitals achieved strong performance in predicting outcomes such as mortality, ICU admission, and mechanical ventilation, but their accuracy often degrades when applied to low-resource settings due to differences in patient populations, disease severity, clinical practices, and documentation quality. Low-resource hospitals also face limited patient volumes, incomplete labeled data, and strict privacy regulations (e.g., HIPAA and GDPR), which prevent centralized data sharing and hinder independent model development, creating a barrier to equitable AI deployment. To address this, we propose a federated transfer learning framework that adapts prognostic models from high-resource to low-resource hospitals without exchanging patient-level data. The approach transfers only aggregate statistics (e.g., feature means, variances, class-conditional distributions, and correlations) via a secure lightweight protocol, enabling target hospitals to align feature distributions using domain adaptation techniques and fine-tune models on small local datasets. The framework includes source model training, statistical aggregation, secure transmission, and target-side adaptation modules, ensuring no raw patient data leaves any institution. By relying on aggregate statistics, the method preserves privacy while mitigating domain shift and maintaining clinical utility across diverse healthcare environments. This scalable and privacy-preserving framework supports broader deployment of COVID-19 predictive models and provides a generalizable strategy for other medical conditions with heterogeneous healthcare settings.