Using hospital Admission, Discharge & Transfer (ADT) data for predicting readmissions

Readmission of patients within a specific period after their discharge from a hospital is a cause of concern for the healthcare industry due to the cost involved. Most of the work done for predicting such readmissions using machine learning (ML) have been based on EHR, claims or authorization data from specific sources, which are mostly snapshot data at one static point in time and hence delayed. ADT being dynamic as the data is available instantaneous on occurrence of a medical event/visit adds value. Our goal is to utilize machine learning on unlabeled ADT data to identify patients who are at a high risk of being readmitted. We approached the problem in three parts. First, we labeled patient events using logical rules and finalized one of many readmission definitions that was more encapsulating of varied scenarios. Second, feature engineering was done which encapsulates the longitudinal timeline of each patient in a representative way considering all the contextual information. Third, we developed an automated machine learning pipeline which takes modeling inputs from the user, runs various models to generate readmission prediction, does a cross validation and returns the best model. We tried multiple combinations of models and cross-validation strategies and decided on a random forest model with specific hyper-parameter values and to be the most effective method to classify high risk patients. It had a test AUC-ROC of 72% which is better than quite a few industry standards. The model currently implemented in the client environment identifies the high-risk patients in real-time to care nurses who in turn take proper interventions to reduce their chances of readmission.