Pursuing Optimal Prediction of Discharge Time in ICUs with Machine Learning Methods

Abstract

In hospital intensive care units (ICU), patients are under continuous evaluation. One of the purposes of this evaluation is to determine the expected number of days to discharge. This value is important to manage ICUs. Some studies show that health care professionals are good at predicting short-term discharge times, but not as good at long-term predictions. Machine learning methods can achieve 1.79-day average prediction error. We performed a study on 3,787 patient-days in the ICU of the Hospital Joan XXIII (Spain) to obtain a data-driven model to predict the discharge time of ICU patients, in a daily basis. Our model, which is based on random forest technology, obtained an error of 1.34 days. We studied the progression of the model as more data are available and predicted that the number of instances required to reduce the error below one day is 4,745. When we trained the model with all the available data, we obtained a mean error of less than half a day with a coefficient of determination (R2) above 97% in their predictions on either ICU survivors and not survivors. Similar results were obtained differentiating by patients’ gender and age, confirming our approach as a good means to achieve optimal performance when more data will be available.