Operationally-Informed Hospital-Wide Discharge Prediction Using Machine Learning

Accurate patient discharge time estimates are invaluable for hospital operations management. They are vital for efficient and effective scheduling of hospital resources including beds and staff. Unexpected discharges place strain on the patient families and care providers, in addition to causing hospital inefficiencies. Due to the increasing availability of electronic health record data, predictive models can be leveraged to not only offer clinical decision support, but also to optimize hospital operations. In this work, we incorporate clinical knowledge from operational leaders at Kaiser Perma-nente Northern California to design a predictive model for patient discharge using a novel dataset that contains hourly data from the electronic health records of 14 different Kaiser Permanente hospitals. We train and test several algorithms with varying complexity to predict patient-level discharges for the following day at operationally relevant times on the hospital-centric timescale. The highest AUC we achieve is 0.729 with a gradient boosted model, which significantly outperforms both the current estimates deployed in these 14 facilities and the baseline model without hourly data. A feature permutation importance assessment is performed and we conclude that the majority of the improvement is due to the inclusion of the detailed, hourly data.