Hospital patients’ length of stay prediction: A federated learning approach

Predicting patient’s length of stay (LOS) is a crucial determinant for hospitals to maintain resource efficiency and quality treatment, where machine learning-based predictive approaches can be extremely beneficial. Though the healthcare industry’s increasing adoption of information technology has transformed it into a massive data hub, the bulk of this data is kept within the medical institution yet not shared with others for confidentiality concerns; which makes it difficult to construct efficient predictive analytics that require vast amount of training data. Hence, this study proposes a federated machine learning-based model for forecasting patients’ LOS in days combining the results from locally trained models of various decentralized and heterogeneous hospital clients while maintaining their data privacy to train an aggregated predictive model in the server. Here, ten hospital client’s administrative data have been trained using three types of machine learning regression models locally utilizing their own data. Then, the parameters (intercept and coefficient) of the locally trained models are sent to the central server in multiple rounds, where they have been aggregated to construct a combined model for LOS prediction. The regression analysis performances of the locally trained models and the server-side model aggregating different number of clients have been compared through various parameter metrics. The findings reveal that, the aggregated model’s predictive performance with federated learning is less error-prone, and that the model’s performance improves when more clients’ parameters are integrated on the server side.