Does comorbidity matrix provide similar amount of predictive information: Comparisons from Charlson and Elixhauser using Deep Learning

Comorbidity information is used in many ways in health outcomes research. The task of finding the best approach to use comorbidity information can be elusive and challenging due to multiple elements of comorbidity information such as flags, scores, combination of flags etc. Charlson and Elixhauser comorbidity indexes were used in this study to answer the following research questions: Do Charlson and Elixhauser scores perform equally well in a deep learning model?; Do Charlson and Elixhauser flags perform equally well in a deep learning model?; Do Charlson and Elixhauser combined flags perform equally well in a deep learning model? These research questions were answered using two types of outcomes (Statin Association Symptoms (SAS) and statin therapy discontinuation). Healthcare claims data from OptumLabs® Data Warehouse (OLDW) was used. There was 9% variation in AUC from our deep learning models predicting SAS, whereas statin therapy discontinuation indicated a difference of 1%. Results indicate that one can gain additional AUC improvement by selecting the best combination of comorbidity information (i.e. scores, flags). Overall, combination of flags produced model with higher AUC indicating an overall better model.