Machine Learning-based Pre-discharge Prediction of Hospital Readmission
Lawrence R. Brindise, Robert JC Steele · 2018
In this paper, we describe the application of data mining techniques in relation to an important clinical care quality indicator: the prediction of hospital readmission within 30 days of discharge. We retrieved six months of encounter data between dates April through September, 2017, from five inpatient hospitals in a large US metropolitan area. Each encounter includes both administrative and clinical data. We utilized a feature reduction technique to replace thousands of clinical features with a much smaller number of proxy features. The dimensionally reduced dataset was then used in the development, training and evaluation of numerous readmission predictive models. Using standard implementation techniques, our model can function within the hospital EHR from where the data was sourced, in real- time, for all patients, prior to discharge. Model performance of the best performing model compares favorably to existing comparable pre-discharge, all-patient predictive model studies.