Development and validation of a machine learning-based claims frailty index for predicting mortality, disability, and healthcare utilization in middle-aged and older adults.

OBJECTIVES: Most existing claims-based frailty indices (CFIs) rely primarily on diagnostic codes, which may overlook functional status reflected in healthcare utilization. We therefore developed and externally validated a machine learning-based CFI integrating diagnoses, medications, and procedures in older adults. METHODS: This cohort study linked the Taiwan Longitudinal Study on Aging (TLSA) with the National Health Insurance Research Database, including participants aged ≥50 years. The TLSA survey-based frailty index (SFI) served as the reference standard. The 2011 cohort (n = 3727) was used for development and the 2015 cohort (n = 5434) for validation. Among 1032 candidate variables, age-associated items were identified and entered into an extreme gradient boosting model (internally validated by 5-fold crossvalidation) to predict the SFI, yielding a 63-item CFI. Associations between the CFI (per 0.1 increment) and outcomes were assessed using Cox models adjusted for age, sex, and comorbidity burden. RESULTS: 723 and 580 deaths occurred in the development and validation cohorts, respectively. Each 0.1-point increase in CFI was associated with higher 1-, 3-, and 5-year mortality (adjusted HR 1.44-1.81, 95% C.I 1.31-2.31), with discrimination comparable to the SFI. Higher CFI scores were also associated with increased risks of healthcare utilization and disability, with discrimination exceeding that of an established multimorbidity frailty index (ΔAUC ranging from 2.79% to 6.90%; P < 0.05). CONCLUSIONS: A machine learning-derived CFI integrating diagnoses, medications, and procedures demonstrated robust prediction of mortality and other adverse outcomes, supporting its incremental value in population-based frailty assessment.