An Interpretable Machine Learning Approach for Predicting Hospital Length of Stay and Readmission

Length of stay (LOS) and risk of readmission of patients are critical indicators of the quality and operation efficiency of hospitals. Various machine learning (ML) approaches have been applied to predict a patient’s hospital LOS and risk of readmission, but those with more accurate predictions are often the so called ‘black-box’ approaches. This study aims to add interpretability in predicting LOS and the risk of readmission in 30-day among all-cause patients admitted through the emergency department (ED) while improving the accuracy and parsimony of the ML approach. Several state-of-the-art ML models were applied to our prediction tasks and their predictive power reported and compared. The CatBoost model outperformed the rest, hence is chosen as the baseline for this study. For interpretability, we introduced Shapley values and analyzed, at both aggregated and individual levels, the prediction results from the CatBoost model. Lower dimension models were further developed following the guidance of Shapley values. Our results show that the lower dimension model can robustly predict hospital LOS and risk of readmission, indicating that Shapley values are not only useful for adding model interpretability, but also effective for creating a lower-dimensional model amenable to implementation.