External validation of model predicting healthcare costs using cohort study-derived and claims-derived predictor variables in community-dwelling older adults

Background: In a pooled analysis of 4 prospective cohort studies linked with Medicare claims (2002–2011), claims-derived predictors and cohort-derived functional impairments and phenotypic frailty were associated with higher subsequent healthcare expenditures in community-dwelling beneficiaries. Objective: Determine calibration and discrimination of published cost prediction models among a nationally representative sample of US community-dwelling older adults in an independent sample (external validation) and at a different time-point (temporal validation). Design: Prospective cohort study Setting: National Healthy Aging and Trends Study (NHATS) 2015 cohort Participants: 3,322 community-dwelling NHATS participants enrolled in Medicare fee-for-service. Measurements: Self-reported functional impairments (difficulty performing four activities of daily living) and frailty phenotype (operationalized using five components) derived from cohort data. Weighted (CMS Hierarchical Conditions Categories index) and unweighted (count of conditions) multimorbidity measures and frailty index derived from claims. Annualized healthcare costs ascertained for 36 months following 2015 examination. Cost ratios estimated from generalized linear models stratified by sex and adjusted for geographic region. Model calibration (predicted vs. observed costs by quintile) and discrimination (area under curve [AUC]) assessed. Results: In external and temporal validation cohorts, calibration was acceptable for all models with substantial agreement between predicted and observed costs. In the external cohort, the full model including claims- and cohort-derived variables had AUC = 0.78, 95% CI (0.74–0.82) in women and AUC = 0.77, 95% CI (0.73–0.81) in men, slightly higher than models limited to only claims-derived variables. Similar findings were noted in the temporal validation cohort. Conclusions: Performance of a model predicting total healthcare costs including self-reported functional impairments and phenotypic frailty together with claims-derived cost indicators is similar in a new independent cohort and in a subset of the original cohort at a later time-point. These results support clinical assessment of these geriatric syndromes for predicting risk of subsequent healthcare burden.