Improving Hospital Readmission Prediction Using Domain Knowledge Based Virtual Examples

In recent years, prediction of 30-day hospital readmission risk received increased interest in the area of Healthcare Predictive Analytics because of high human and financial impact. However, lack of data, high class and feature imbalance, and sparsity of the data make this task so challenging that most of the efforts to produce accurate data-driven readmission predictive models failed. We address these problems by proposing a novel method for generation of virtual examples that exploits synergetic effect of data driven models and domain knowledge by integrating qualitative knowledge and available data as complementary information sources. Domain knowledge, presented in the form of ICD-9 hierarchy of diagnoses, is used to characterize rare or unseen co-morbidities, which presumably have similar outcome according to ICD-9 hierarchy. We evaluate the proposed method on 66,994 pediatric hospital discharge records from California, State Inpatient Databases (SID), Healthcare Cost and Utilization Project (HCUP) in the period from 2009 to 2011, and show improved prediction of 30-day hospital readmission accuracy compared to state-of-the-art alternative methods. We attribute the improvement obtained by the proposed method to the fact that rare diseases have high percentage of readmission, and models based entirely on data usually fail to detect this qualitative information.