Machine-Learned Comorbidity Index Improves Patient Risk Assessment

Traditional comorbidity scores, such as the Charlson and Elixhauser indices, are widely used in healthcare for risk adjustment and patient stratification. However, these conventional scores suffer from two primary limitations. Firstly, they are largely focused on mortality outcomes, which means they may not accurately align with other crucial clinical outcomes. Secondly, their linear, rule-based structure prevents them from capturing the complex, nonlinear, and outcome-specific risk relationships that exist within patient data. To address these shortcomings, researchers have proposed a novel Machine-Learned Comorbidity Index (MLCI). This index is designed to map diagnosis codes to a single scalar value. Its development involves maximizing the normalized Hilbert-Schmidt Independence Criterion (nHSIC) between the learned score and multiple clinical outcomes. This approach allows MLCI to effectively capture nonlinear dependencies between risk factors and patient outcomes. The theoretical underpinnings of MLCI also characterize the conditions under which a unified, informative admission-level ordering can be achieved across various outcomes. Empirical evaluations using multiple benchmark electronic health record (EHR) datasets demonstrated that MLCI consistently outperforms strong baseline models across several evaluation metrics, indicating its potential to provide a more accurate and comprehensive assessment of patient risk and comorbidity.