Keywords: Hospital Readmissions, Support Vector Machine, Random Forest, Decision Tree, Logistic Regression, Neural Network, Predictive Model, Health Care Quality
CDPHP seeks to prevent 30 day all-cause hospital readmissions to improve quality of care and reduce costs. The primary strategy is the use of case managers who assist members with through educational outreach and communication with their providers. Proactively suggesting case management services to every member with a non-acute admission is an inefficient use of resources, so a predictive model was built to target members with a high probability of a readmission. Five statistical and machine learning models were tested, and each was evaluated for computational expense and predictive accuracy. The models identified relative risk score, history of inpatient and ED visits, and primary admitting diagnosis as significant factors. The logistic regression model had the strongest performance across the desired conditions and had a c-statistic of 0.67, sensitivity of 67 percent and specificity of 68 percent. The model is run daily to assign a predicted probability of readmission for members with a prior authorization for a non-acute inpatient admission. The result is an efficient and effective way for case managers to prioritize outreach to members to prevent subsequent readmissions.