A Smart Healthcare Management System for Predicting ICU Length of Stay Using Machine Learning and MIMIC-IV

Stays in ICUs are expensive to accommodate and difficult to forecast, yet even a simple forecast can be helpful, as it enables ICUs to allocate beds and staff and schedule transfer times. We address a particular problem of this kind: how much of the length of stay can be determined based on the first day of hospitalization? Our dataset is the publicly available MIMIC-IV Clinical Database Demo, from which we use ۱۴۰ hospitalization records, retain the ۱۱۷ that lasted at least one day, and derive ۴۷ features from the first ۲۴ hours of data on vital signs, medication and fluid intake, urine volume, and bedside procedures. Both formulations are addressed for the same set of patients. Exact forecasting of duration turned out to be rather complicated in our case: despite selecting the best-performing regression technique (random forest), the resulting mean absolute error was ۲.۸۵ days, barely outperforming the mean baseline predictor and producing a negative value for the coefficient of determination due to a few very long stays contributing disproportionately to the squared error. Yet, what has been accomplished is the ranking, where the Spearman rank correlation.

کلیدواژه ها:

ICU Length of Stay ، Machine learning ، Electronic Health Records ، MIMIC-IV ، Predictive Modelling ، Critical Care