New outcome-specific comorbidity scores excelled in predicting in-hospital mortality and healthcare charges in administrative databases

To determine the most reliable comorbidity measure, we adapted and validated outcome-specific comorbidity scores to predict mortality and hospital charges using the comorbidities composing the Charlson and Elixhauser measures and the combination of these two used in developing Gagne's combined comorbidity scores (CC, EC, and GC, respectively). We divided cases of patients discharged in 2016–17 from the Diagnosis Procedure Combination database (n = 2,671,749) into two: one to derive weights for the scores, and the other for validation. We further validated them in subgroups, such as that with a selected diagnosis. The c-statistics of the models predicting in-hospital mortality using new mortality scores using the CC, EC, and GC were 0.780, 0.795, and 0.794, respectively. Among them, that using the EC showed the best calibration. To predict hospital charges and the length of hospital stay (LOS), the models using variables indicating the GC performed the best. The performances of the mortality and expenditure scores were considerably different in predicting each outcome. The new score using the EC performed the best in predicting in-hospital mortality for most situations. For hospital charges and the LOS, the binary variables of the GC showed the best results. The outcome-specific comorbidity scores should be considered for different outcomes.