Improving 30-day readmission risk predictions using machine learning

Hospital readmission is used as a proxy measure of effectiveness of care providers in New Zealand and overseas, and risk predictions can be used to identify at-risk patients in order to reduce preventable readmissions. Existing objective measures for measuring readmission risk potentially sacrifice accuracy in order to produce models that are simple. This study investigates the hypothesis that using machine learning on a wide feature set will improve the accuracy of readmission risk predictions compared with existing techniques. A publicly available dataset containing 100,000 admissions and 56 features was used to evaluate the hypothesis. The results were encouraging: although the best results were achieved using logistic regression – the same machine learning algorithm used successfully in existing measures – the model improved significantly over the measures derived from a simpler feature set represented in the LACE tool.