Pre-assessment of machine learning approaches for patient length of stay prediction

Abstract

Patient length of stay (LOS) in ICU and hospital’s general care unit is one of the important indicators that is widely measured and benchmarked to improve the quality and efficiency of patient care. There are many studies both on statistical testing of the LOS outcome to determine factors associated with it and the outcome predictive modeling using machine learning algorithms. However, there are still fewer studies of the LOS outcome predictive modeling using local datasets. Therefore, an initial study of assessing supervised machine learning approaches for regression and classification tasks on predicting the LOS outcome has been conducted using an aggregated Multiparameter Intelligent Monitoring in Intensive Care-III (MIMIC-III) public dataset to obtain an overview of the prediction framework and to compare the predictive performance of the machine learning models. This is as a preparation for developing an outcome calculator of the LOS outcome and other operative outcomes in patients based on local data. The LOS was predicted using 10 classification and 15 regression models assessed using accuracy, precision, recall, and F1 scores for the classification task and root mean squared error (RMSE) for the regression task. The results showed that the Extreme Gradient Boosting classifier and Extreme Gradient Boosting regressor presented the highest validation and good testing performances for the LOS prediction. Although model overfitting trend was detected in both models’ learning curves, in this respect, the study may serve as a useful starting point for an extended work associating with predictive modeling on the LOS and other operational outcomes.