Predicting population health focused outcomes using machine learning

Care management activities seek to reduce healthcare cost and improve patient outcomes. Identifying patients who may receive substantial benefit from care management services can be especially challenging when managing large populations across disparate systems. This research tests a novel method for identifying patients for care management using over 30 disparate healthcare data sources and machine learning. Random Forest models were used to predict four binary outcomes; high cost, hospital admission, hospital readmission, and multiple emergency department visits. The models leveraged population health enterprise data warehouse cross-ontology mappings for the following data types; conditions, procedures, medications, results, demographics, and claims-based cost and utilization. Each of the data types were tested independently then combined incrementally. The highest performing models for each outcome of interest resulted with the following ROC AUC; High Cost (0.81), Admission (0.80), Re-admission (0.86), and Multi-ED (0.74). The research shows disparate data sources and machine learning can be used to predict population health focused outcomes. The framework used in this research has the potential to expand and scale to include any number of additional data types and outcomes.Ph.D.Includes bibliographical reference