Hospital Readmission Prediction using Discriminative patterns

Avoidable hospital readmission is problematic as it increases the burden on healthcare systems, leads to a shortage of hospital beds and impacts on the costs of healthcare. Various machine learning algorithms have been applied to predict patient readmissions. However, existing literature has only focused on individual features of health conditions without consideration of associations between features. This paper proposes discriminative pattern-based features as a technique to improve readmission prediction. First, discriminative patterns that occur disproportionately between two classes: readmission and non-readmission, were generated based on hospital electronic health records. Second, the patterns were fed as features into a classification model for readmission prediction. Then, we have evaluated these discriminative pattern-based features in three datasets: diabetes, chronic kidney disease and all diseases. Experiments with each dataset showed that the features of chronic disease cohorts have fewer differences between the readmission and the non-readmission classes than the all-diseases cohort. Our proposed pattern-based model improved the prediction performance in terms of AUC (Area Under the receiver operating characteristic curve) by about 12% compared with the baseline models for the all-disease cohort, however, it showed little improvement for either diabetes or chronic kidney disease datasets.