Predicting 120-day hospital readmission using medical administrative patient data

Hospitals and health-care insurers routinely use models to predict patient readmission, extrapolating from historical data. Subsequently, the predicted quantities can be used for contracting and pricing negotiations between these hospitals and healthcare insurers. The Dutch healthcare system uses unique standardized Care Trajectories (so-called DBCs) for administration and billing of care. Here, we compared supervised machine learning methods on predicting 120-day readmission as an operationally significant metric. We used administrative patient data from 21 common Care Trajectories, in combination with demographic information. A lightGBM model using undersampling to tackle class imbalance yielded an AUROC score of 0.86 and provided the highest recall score (73.8%).