2017 HSR&D/QUERI National Conference
1050 — Developing an Improved Comorbidity Summary Score through Modeling ICD Data from Electronic Health Records Data
Lead/Presenter: Ralph Ward, COIN - Charleston
All Authors: Ward RC (VA Charleston COIN)
Gebregziabher (VA Charleston COIN)
Ramakrishnan V (Medical University of South Carolina)
Axon N (VA Charleston COIN)
Frey L (VA Charleston COIN)
Egede L (Medical College of Wisconsin)
Accurate adjustment for comorbidities (or disease burden) is essential in healthcare disparities research in order to minimize the risk of bias. Commonly-used adjustment methods (Charlson or Elixhauser comorbidity indices) often provide poor results. In this study, we develop an improved comorbidity summary score and compare its performance to the Elixhauser index.
We propose a measure of disease burden developed by comparing seven machine learning and maximum likelihood methods (including random forest, Bayesian additive regression trees, elastic-net regression) for dimension reduction in determining a latent comorbidity score. The operating characteristics of the proposed measure are examined using receiver operating characteristics (ROC) curve statistics and cross-validation techniques.
We applied this approach to retrospective data involving three large cohorts of Veterans, including 892,223 with diabetes, 3,359,560 with chronic kidney disease and 168,521 with a history of traumatic brain injury. We showed that substantially improved risk adjustment over existing comorbidity indices can be achieved through a summary measure derived through the predictive modeling techniques demonstrated here. For example, mean area under the ROC curve (AUC) statistics and 95% CI for the DM, CKD and TBI cohorts were 0.84 (0.83, 0.85), 0.81 (0.80, 0.82), 0.84 (0.83, 0.85) respectively, for models based on the new summary score, compared with 0.72 (0.71, 0.74), 0.72 (0.70, 0.73) and 0.78 (0.77, 0.80), respectively, for models based on the Elixhauser comorbidity index.
The proposed comorbidity score based on the combined strengths of machine learning algorithms and traditional statistical methods produced more robust adjustment for disease burden.
Researchers should consider alternative methods to previously validated comorbidity scores when adjusting for risk in health services and outcomes research. Such gains in risk prediction modeling can have a wide impact across numerous studies that attempt to understand disparate healthcare outcomes between various Veteran groups.