Abstract
Background/Aims Health organizations are beginning to apply predictive analytics as a central and critical tool for more effective healthcare management. However, the art is still far from maturity, and it is necessary to develop and perfect the requisite analytic tools. A need exists for methods to measure illness burden and identify patients for targeted interventions. Most commercial programs are unable to use all of the data we have available for analysis. Their input is limited to age, gender, diagnoses and medications, while our database also contains a wide range of demographic, socioeconomic, clinical and financial data at the patient level. We hypothesized that utilizing the richer data would generate robust analytic and predictive capabilities. We then developed a predictive analytical system that accesses our entire database. The design requirements included flexible and generic database mapping and transparency of any algorithm’s internal processes. In addition, the system has embedded quality assurance processes and maintains an historical record of all analytical models and results.
Methods Data sources included approximately 15 years of history of physician and other medical professional visits, hospitalizations, emergency room visits, diagnoses, medications, laboratory results, imaging studies, pathology results, and extensive socio-economic, demographic data and associated costs of all medical expenditures. Analytics techniques used included linear regression, classification trees, and additional data mining methods. Models developed included: predicted annual cost and prediction of re-hospitalization within 30 days. Models were validated using R2, C-statistics and Positive Predictive Value (PPV).
Results The first model (R2 ~ 0.36) was used to create reports for risk adjustment and phyician profiling. The second model (PPV 54%) was incorporated into an existing program for preventing re-hospitalization.
Conclusions The Maccabi analytical tool has a robust predictive ability and has been successfully used for physician profiling and predicting re-hospitalization. We suggest evaluating this tool on different databases to yield insight into its transferability and robustness. The minimal required data set for use in other organizations needs to be determined.
