An Open Algorithm for Systematic Evaluation of Readmission Predictors on Diabetic Patients from Data Warehouses

摘要

Prediction models depend on the variables or input attributes that characterize the studied phenomenon. When data collected from databases and the phenomenon to modeling is complex (e.g. patient's health prediction), in many cases there is a need to construct representative variables according to the problem and then modeling and testing. Both processes can be automatized using algorithms to assess their fit with real data. A proposed algorithm wraps this modeling and testing process and looks for the variables constructed from knowledge, helping to fit to real data, and which could be used for any modeling problem that has both quantitative and qualitative mixed information. To test this approach, real data from ten years of diabetic patients records was used following the American Diabetes Association last recommendations to construct variables that could find high risk patients and evaluate the variables efficacy. The method could be used to improve data warehouse's framework and for this case, to help care institutions to deploy new health politics to adjust treatments and resource management for diabetic patients.