Development of risk prediction models for incident frailty and their performance evaluation

There is currently no tool to predict incident frailty despite various frailty assessment tools. This study aimed to develop risk prediction models for incident frailty and evaluated their performance on discrimination, calibration, and internal validity. This 2-year follow-up study used data from 5076 non-frail older adults (51% women) living in Tokyo at baseline. We used the Kaigo-Yobo checklist, a standardised assessment instrument, to determine frailty. Twenty questionnaire-based variables that include sociodemographic, medical, behavioural, and subjective factors were entered into binary logistic regression analysis with stepwise backward elimination (p < 0.1 for retention in the model). Discrimination and calibration were assessed by area under the receiver operating characteristic curve (AUC) and the Hosmer-Lemeshow test, respectively. For the assessment of internal validity, we used a 5-fold cross-validation method and calculated the mean AUC. At the follow-up survey, 15.0% of men and 10.2% of women were frail. The frailty risk prediction model was composed of 10 variables for men and 11 for women. AUC of the model was 0.71 in men and 0.72 in women. The P-value for the Hosmer-Lemeshow test in both models was more than 0.05. For internal validity, the mean AUC was 0.71 in men and 0.72 in women. Probability of incident frailty rose with an increasing risk score that was calculated from the developed models. These results demonstrated that the developed models enable the identification of non-frail older adults at high risk of incident frailty, which could help to implement preventive approaches in community settings.