Using machine learning to improve readmission risk in surgical patients in South Africa

Abstract

Unplanned readmission within 30 days is a major challenge both globally and in South Africa. The aim of this study was to develop a machine learning model to predict unplanned surgical and trauma readmission to a public hospital in South Africa from unstructured text data. A retrospective cohort of records of patients was subjected to random forest analysis, using natural language processing and sentiment analysis to deal with data in free text in an electronic registry. Our findings were within the range of global studies, with reported AUC values between 0.54 and 0.92. For trauma unplanned readmissions, the discharge plan score was the most important predictor in the model, and for surgical unplanned readmissions, the problem score was the most important predictor in the model. The use of machine learning and natural language processing improved the accuracy of predicting readmissions.