Predicting Patient Behaviour in Swedish Health Care Using Machine Learning

Abstract

There is a vast amount of patient data stored in health care record systems. Together with the rise of computing power this data could be used for advanced analysis of this data, and incorporate it in applications for use in daily operations. This is a case study in which unbalanced archival data from emergency room admissions is used for machine learning, in order to develop three models that predict the possibility of a patient returning to emergency room within 72 hours. The best of these model uses a logistic regression classifier and has a recall of 1% and a precision of 50%. The implementation of such a model in daily operation is discussed with a new approach to cost benefits. Despite the low predictability, the study is a proof of concept of predictive modeling in a health care context.