Counting co-occurring diseases to predict mortality is as accurate as multimorbidity indices: an external validation study
Source: PubMed Central Open Access, NCBI / U.S. National Library of Medicine
Abstract Background A systematic review recommended seven multimorbidity indices for predicting mortality. However, their performance has not been assessed in a head-to-head comparison. We externally validated these indices and determined their performance compared to counting co-occurring diseases. sec3 Setting Within the prospective Rotterdam Study in the Netherlands, we constructed seven specific sub-cohorts, selected from 14 926 community-dwelling older adults to match the target population of the selected multimorbidity indices. sec4 Methods We calculated prediction scores according to the indices’ original methods and used these as predictors in logistic regression models with all-cause mortality as outcome. We assessed their performance and compared it to four benchmark models fitted on the same index-specific samples. These models were based on (i) age and sex; (ii) counts of co-occurring diseases, age and sex; (iii) counts of co-occurring diseases associated with mortality, age and sex; and (iv) individual diseases as separate predictors, age and sex. sec5 Results The total population sizes of the seven sub-cohorts ranged from 2409 to 9045 participants. The mean age of the populations ranged from 59.4 to 77.0 years; the proportion of women ranged from 56.0% to 61.8% (excluding single-sex indices). The absolute risk for mortality ranged from 0.9% to 13%. Discriminative performance of the indices and corresponding count models was nearly identical across all indices (m
Abstract
Abstract Background A systematic review recommended seven multimorbidity indices for predicting mortality. However, their performance has not been assessed in a head-to-head comparison. We externally validated these indices and determined their performance compared to counting co-occurring diseases. sec3 Setting Within the prospective Rotterdam Study in the Netherlands, we constructed seven specific sub-cohorts, selected from 14 926 community-dwelling older adults to match the target population of the selected multimorbidity indices. sec4 Methods We calculated prediction scores according to the indices’ original methods and used these as predictors in logistic regression models with all-cause mortality as outcome. We assessed their performance and compared it to four benchmark models fitted on the same index-specific samples. These models were based on (i) age and sex; (ii) counts of co-occurring diseases, age and sex; (iii) counts of co-occurring diseases associated with mortality, age and sex; and (iv) individual diseases as separate predictors, age and sex. sec5 Results The total population sizes of the seven sub-cohorts ranged from 2409 to 9045 participants. The mean age of the populations ranged from 59.4 to 77.0 years; the proportion of women ranged from 56.0% to 61.8% (excluding single-sex indices). The absolute risk for mortality ranged from 0.9% to 13%. Discriminative performance of the indices and corresponding count models was nearly identical across all indices (maximum difference in C-statistic: 0.06), yet higher than age-and-sex models. Absolute accuracy of the prediction scores was similar across all models (maximum improvement in Brier score: 4%). Calibration was poor in four out of seven indices, all of which had a follow-up time of 2 years or less. sec6 Conclusion Counting co-occurring diseases is as accurate in predicting all-cause mortality in the general population as using multimorbidity indices. These findings imply that counting diseases is the more practical and reliable way of providing prognosis to patients with multimorbidity in a population of community-dwelling adults. sec7