Development and validation of a machine learning model for frailty screening using claims data in Japan: the Longevity Improvement & Fair Evidence Study.

The XGBoost-based claims data model achieved an ROC-AUC of 0.780 in internal validation and 0.728 in external validation for predicting frailty status.

• Model was trained using the eXtreme Gradient Boosting (XGBoost) algorithm
• Phase 1 included 74,148 individuals total: development cohort of 60,930 and validation cohort of 13,218 from a single municipality
• Internal validation ROC-AUC: 0.780; external validation ROC-AUC: 0.728
• Model incorporated demographic variables, long-term care use, comorbidities, procedures, and medical device use

Frailty classification derived from the claims-based model was associated with significantly higher all-cause mortality in both development and validation cohorts.

• Phase 2 external validation was conducted in a new cohort of 354,815 individuals from seven other municipalities
• Hazard ratio for mortality in the development cohort: 7.03 (95% CI: 6.47–7.63)
• Hazard ratio for mortality in the validation cohort: 6.75 (95% CI: 6.62–6.89)
• Mortality risk was estimated using the Kaplan-Meier method and Cox regression models

The frailty prediction model was developed as a scalable alternative to the resource-intensive Questionnaire for Medical Checkup of Old-Old (QMCOO) used in Japan.

• The QMCOO is the standard tool used in Japan to assess frailty but is described as resource-intensive to implement
• The model used administrative claims data, which are routinely collected, to predict QMCOO-defined frailty status
• The study was conducted under the Longevity Improvement & Fair Evidence (LIFE) Study framework
• The model is intended to support population-level frailty screening where questionnaire-based assessments are impractical

The study used a two-phase validation design to assess both predictive accuracy and prognostic utility of the frailty model.

• Phase 1 assessed model discrimination for frailty classification using data from a single municipality
• Phase 2 evaluated the model's prognostic utility for predicting all-cause mortality using data from seven other municipalities
• Model performance was assessed primarily using the area under the receiver operating characteristic curve (ROC-AUC)
• The two-phase design allowed separation of classification performance from clinical prognostic value