Development and validation of the two morbidity indices for health-state severity adjustment, using data from the French National Health Insurance Information

Abstract

Introduction Morbidity indices are commonly used for adjustment or stratification on health-state severity, for research purposes and to inform policy-makers. The performance of morbidity indices is enhanced when multiple data sources are combined to identify conditions included as predictors and when indices are used for the same outcome and in similar settings as for which they were developed. Most existing indices use a single data source to identify conditions and have been developed to predict a single outcome or subsequently adapted to heterogeneous settings. The French National Health Insurance Information System (“Systeme National des Donnees de Sante, SNDS”) provides individual expenditure and morbidity information for each beneficiary of the main health insurance scheme, representing over 57 million individuals (87% of the French population). Morbidity is measured through a set of standard algorithms combining ICD-10 diagnoses and pharmacy data to identify specific conditions. We used SNDS data to develop and validate two outcome-specific morbidity indices: the Mortality-Related Morbidity Index (MRMI) predictive of all-cause mortality and the Expenditure-Related Morbidity Index (ERMI) predictive of healthcare expenditure. Methods A cohort including all beneficiaries of the main French health insurance scheme aged 65 years or older on December 31, 2013 (N = 7,672,111) was randomly split into a development population for index elaboration and a validation population for predictive performance assessment. Age, gender and selected lists of conditions were used as predictors for 2-year mortality and 2-year healthcare expenditure in separate models. Predictors were selected according to face validity of their identifying algorithm, correlation analysis within disease categories, univariate association with each outcome and outcome-specific considerations: number of events for mortality prediction and variance explained for expenditure prediction. To derive a weighted index from the adjusted regression coefficients, we applied a scoring rule where each additional point reflected risk associated with a 5-years age increase. We conducted two series of sensitivity analyses to test the stability of the weights when accounting for effect modification due to age and gender or to associations of conditions. Overall performance and calibration of the MRMI and ERMI were measured and compared to various versions of the Charlson index (CCI). For the MRMI, we assessed discrimination using the concordance statistic (c-statistic, equivalent to the area under the receiver operating characteristic curve). For the ERMI, we measured the percentage of total expenditure variance explained by the index. Calibration was assessed, for both indices, by comparing predictions with observations among individuals with the same index value. Results The MRMI included 16 conditions weighted 1 to 3. Adjusted odds ratios [95% confidence interval] ranged from 1.34 [1.31–1.37] for depression, 1.36 [1.32–1.39] for Rheumatic or connective tissue diseases, to 3.79 [3.62–3.97] for end-stage renal disease, 4.09 [4.04–4.15] for cancer. The MRMI was more discriminant than the age-adjusted CCI (c-statistic: 0.8252 [95% CI: 0.8244–0.8260] vs. 0.7999 [0.7991–0.8008]) and better calibrated. Age and gender alone (c-statistic: 0.7613 [0.7604–0.7623]) were more discriminant than indices that did not include age (original CCI and cost-adapted CCI). The ERMI included 19 conditions weighted 2 to 16, explained more variance than the cost-adapted CCI (21.8% vs. 13.0%) and was better calibrated. Sensitivity analyses showed that the estimates and resulting weights were stable for both indices. Conclusions The MRMI and ERMI indices are performant tools to account for health-state severity according to outcomes under study. They also document the differential adjusted effect of conditions on different outcomes.