Predicting 30-day all-cause readmissions from hospital inpatient discharge data

Inpatient hospital readmissions for potentially avoidable conditions are problematic and costly. In this paper, we build machine learning models using variables widely available in health claims data to predict patients' 30-day readmission risks at the time of discharge. These models show high predictive power on a U.S. nationwide readmission database. They are also capable of providing interpretable risk factors globally at the population level and locally associated with each single discharge. In addition, we propose a model-agnostic approach to provide confidence for each prediction. Altogether, using models with high predictive power, interpretable risk factors and prediction confidence may enable health care systems to accurately target high-risk patients and prevent recurrent readmissions by accurately anticipating the probability of readmission at the point of care.