Publications

Appropriate nurse staffing is a critical challenge for healthcare delivery institutions, that can affect the delivery of quality patient care. Reliable forecasting of nursing demand can optimize resource utilization and support preemptive hiring strategies; however, the use of advanced machine learning approaches to address this challenge has been limited thus far. We propose a probabilistic forecasting approach for predicting nursing workforce demand across multiple hospital units using DeepAR, an autoregressive recurrent neural network algorithm. We analyzed 5-year historical workforce data from a large US-based health system in New York, encompassing both full-time and temporary nurses across multiple specialties and hospital units. Our implementation leverages DeepAR's ability to create a single global model for heterogeneous time-series, incorporating the impact of the COVID-19 pandemic, through binary feature encoding. The model was trained on 38 business unit-specialty combinations, each containing over 100 employees, with a 12-month context and prediction length. The model successfully captured complex temporal patterns, with most ground truth values falling within the 90% prediction intervals and accurately predicting both gradual trends as well as sudden magnitude changes. The model achieved consistent performance across multiple test sets through rolling window analysis, with only 9 out of 38 combinations showing significant temporal sensitivity. Furthermore, validation on 91 smaller business unit-specialty combinations (with 15-100 employees) showed no statistically significant performance degradation, indicating strong generalizability. This approach provides healthcare administrators with a reliable tool for probabilistic workforce demand forecasting, enabling more informed staffing decisions across diverse hospital settings.

Objective: To develop a machine learning prediction model for ambulatory appointment non-arrivals that can be deployed across multiple medical specialties. Methods: We analyzed 4.3 million ambulatory appointments from 1.2 million adult patients using the XGBoost machine learning algorithm. The model incorporated patient demographics, appointment history, provider information, weather data, and lead time. Results: The XGBoost model achieved the highest predictive performance (AUC 0.768). The most important features included rescheduled appointments, lead time, appointment provider, days since last appointment, and prior appointment status. The model calibrated well across all departments, especially for the operationally relevant 0-40% non-arrival probability range. Clinical Application: The model can be integrated into electronic health systems or dashboards to identify high-risk patients and reduce no-shows.