Automate Creating, Customizing, and Optimizing Comorbidity Indices Using a Data-Driven AI/ML Approach

Due to individual differences in severity of illness, clinical studies typically use a comorbidity index to adjust outcomes. With the increasing use of electronic health records (EHRs) to assess the quality of care, a key question arises: how to adjust outcomes and control for severity effectively. Although one may adjust outcomes by using an existing comorbidity index, a suboptimal adjustment problem may occur when applying the existing comorbidity index to different outcomes or patient subgroups. Researchers can develop a new comorbidity index or modify an existing one to address this issue, but it requires time. This study proposes an Automatically Customized Comorbidity Index (ACCI) algorithm to automatically create, customize, and optimize a comorbidity index with EHR and a user's outcome of interest via prediction and optimization components. As an example, we use ACCI to create comorbidity indices for three outcomes of interest: statin-associated symptoms, statin therapy discontinuation, and statin days-supply. Here, we use random forest as the prediction and genetic algorithm as the optimization components. The result shows that ACCI iteratively improved the comorbidity index's prediction and relevance to the outcome. Those comorbidity indices also outperform the baselines, Charlson and Elixhauser comorbidity indices.