1.Validation and Recalibration of Charlson and Elixhauser Comorbidity Indices to Predict In-hospital Mortality in Hospitalized Patients in a Japanese Hospital-Based Administrative Database
Tomomi KIMURA ; Toshifumi SUGITANI ; Takuya NISHIMURA ; Masanori ITO
Japanese Journal of Pharmacoepidemiology 2020;25(1):1-14
Objective: To validate and recalibrate Charlson and Elixhauser comorbidity indices (CCI and ECI, respectively) in a Japanese hospital-based administrative database.Methods: In this retrospective, cohort study, derivation and validation cohorts were developed to include all hospitalizations for patients aged ≥ 18 years at admission and discharged in 2015 or 2016, respectively, from an administrative database based on 287 hospitals. Seventeen CCI and 30 ECI conditions were identified using the International Classification of Diseases (ICD) -10 codes at admission or during the stay. Predictability for hospital death was evaluated using C statistics from multivariable logistic regression models including age, sex, and individual CCI/ECI conditions or the CCI/ECI score in the derivation cohort. After stepwise selection, weighted risk scores were re-assigned to each condition based on the odds ratios (CCI) or beta-coefficient (ECI), and these modified models were evaluated in the validation cohort.Results: The original CCI/ECI had good predictive abilities for hospital death: C statistics (95% confidence interval) for individual comorbidities and score models were 0.764 (0.762-0.765) and 0.731 (0.729-0.733) for CCI, and 0.783 (0.781-0.784) and 0.750 (0.748-0.752) for ECI, respectively. Modified CCI and ECI had 13 and 27 conditions, respectively, but maintained comparable predictive abilities: C statistics for modified individual comorbidities and score models were 0.761 (0.759-0.763) and 0.759 (0.757-0.760) for CCI, and 0.784 (0.782-0.785) and 0.783 (0.781-0.785) for ECI, respectively.Conclusions: The original and modified CCI/ECI models, with reduced numbers of conditions, had sufficient and comparable predictive abilities for hospital death and can be used in future studies using this administrative database.
2.Validation and Recalibration of Charlson and Elixhauser Comorbidity Indices Based on Data From a Japanese Insurance Claims Database
Tomomi KIMURA ; Toshifumi SUGITANI ; Takuya NISHIMURA ; Masanori ITO
Japanese Journal of Pharmacoepidemiology 2019;24(2):53-64
Objective: The Charlson and Elixhauser comorbidity indices (CCI and ECI, respectively) are widely used to study comorbid conditions but these indices have not been validated in Japanese datasets. In this study, our objective was to validate and recalibrate CCI and ECI in a Japanese insurance claims database.Methods: All hospitalizations for patients aged≥18 years discharged between January 2011 and December 2016 were randomly allocated to derivation and validation cohorts. Predictability for hospital death and re-admission was evaluated using C statistics from multivariable logistic regression models including age, sex, and individual CCI/ECI conditions at admission month or the derived score in the derivation cohort. After stepwise variable selection, weighted risk scores for each condition were re-assigned using odds ratios (CCI) or beta coefficients (ECI). The modified models were evaluated in the validation cohort.Results: The original CCI/ECI had good discriminatory power for hospital death: C statistics (95% confidence interval) for individual comorbidities and score models were 0.845 (0.835-0.855) and 0.823 (0.813-0.834) for CCI, and 0.839 (0.828-0.850) and 0.801 (0.790-0.812) for ECI, respectively. Modified CCI and ECI had reduced numbers of comorbidities (17 to 10 and 30 to 21, respectively) but maintained comparable discriminatory abilities: C statistics for modified individual comorbidities and score models were 0.843 (0.833-0.854) and 0.838 (0.827-0.848) for CCI, and 0.840 (0.828-0.852) and 0.839 (0.827-0.851) for ECI, respectively.Conclusions: The original and modified models showed comparable discriminatory abilities and both can be used in future studies using insurance claims databases.