Imbalanced Learning for Hospital Readmission Prediction using National Readmission Database

In this paper, we propose to use imbalanced learning for hospital readmission prediction. The goal is to predict whether a patient, based on his/her current hospital visit records, is likely going to be re-admitted or not within 30-days after being discharged from the current hospital visit. The main challenge of hospital readmission prediction is twofold: (1) the readmission visits (i.e., the positive class) are a small portion of the total hospital visits, representing a severe class imbalance problem for learning; (2) due to privacy and health regulation, the information available for patient characterization is limited; and is often only limited to the payment level information. However, there are over 80,000 procedures code, representing a high dimensionality and high sparsity problem for learning. Motivated by the above challenges, in this paper, we design an imbalanced learning strategy to create features from patient hospital visit, by combining patient demographic information, ICD-10 clinical modification (CM) and procedure codes (PCS), and Clinical Classification Software Refined (CCSR) conversion. Instead of directly using ICD-10-CM/PCS code to characterize patients, we convert each patient’s visit to CCSR code space with a smaller feature space. By using random sampling approach to balance the sample distributions in the training set, our method achieves good performance to predict patient readmission.