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

Development and validation of a preliminary clinical support system for measuring the probability of incident 2-year (pre)frailty among community-dwelling older adults: A prospective cohort study

OBJECTIVE

To develop the wed-based system for predicting risk of (pre)frailty among community-dwelling older adults.

MATERIALS AND METHODS

(Pre)frailty was determined by physical frailty phenotype scale. A total of 2802 robust older adults aged ≥60 years from the China Health and Retirement Longitudinal Study (CHARLS) 2013-2015 survey were randomly assigned to derivation or internal validation cohort at a ratio of 8:2. Logistic regression, Random Forest, Support Vector Machine and eXtreme Gradient Boosting (XGBoost) were used to construct (pre)frailty prediction models. The Grid search and 5-fold cross validation were combined to find the optimal parameters. All models were evaluated externally using the temporal validation method via the CHARLS 2011-2013 survey. The (pre)frailty predictive system was web-based and built upon representational state transfer application program interfaces.

RESULTS

The incidence of (pre)frailty was 34.2 % in derivation cohort, 34.8 % in internal validation cohort, and 32.4 % in external validation cohort. The XGBoost model achieved better prediction performance in derivation and internal validation cohorts, and all models had similar performance in external validation cohort. For internal validation cohort, XGBoost model showed acceptable discrimination (AUC: 0.701, 95 % CI: [0.655-0.746]), calibration (p-value of Hosmer-Lemeshow test > 0.05; good agreement on calibration plot), overall performance (Brier score: 0.200), and clinical usefulness (decision curve analysis: more net benefit than default strategies within the threshold of 0.15-0.80). The top 3 of 14 important predictors generally available in community were age, waist circumference and cognitive function. We embedded XGBoost model into the server and this (pre)frailty predictive system is accessible at http://www.frailtyprediction.com.cn. A nomogram was also conducted to enhance the practical use.

CONCLUSIONS

A user-friendly web-based system was developed with good performance to assist healthcare providers to measure the probability of being (pre)frail among community-dwelling older adults in the next two years, facilitating the early identification of high-risk population of (pre)frailty. Further research is needed to validate this preliminary system across more controlled cohorts.

Frailty Risk Prediction Model among Older Adults: A Chinese Nation-Wide Cross-Sectional Study

OBJECTIVES

Numerous studies have been performed on frailty, but rarely do studies explore the integrated impact of socio-demographic, behavioural and social support factors on frailty. This study aims to establish a comprehensive frailty risk prediction model including multiple risk factors.

METHODS

The 2018 wave of the Chinese Longevity and Health Longitudinal Survey was used. Univariate and multivariate logistic regressions were performed to identify the relationship between frailty and multiple risk factors and establish the frailty risk prediction model. A nomogram was utilized to illustrate the prediction model. The area under the receiver operating characteristic curve (AUC), Hosmer-Lemeshow test and calibration curve were used to appraise the prediction model.

RESULTS

Variables from socio-demographic, social support and behavioural dimensions were included in the final frailty risk prediction model. Risk factors include older age, working as professionals and technicians before 60 years old, poor economic condition and poor oral hygiene. Protective factors include eating rice as a staple food, regular exercise, having a spouse as the first person to share thoughts with, doing physical examination once a year and not needing a caregiver when ill. The AUC (0.881), Hosmer-Lemeshow test (p = 0.618), and calibration curve showed that the risk prediction model was valid.

CONCLUSION

Risk factors from socio-demographic, behavioural and social support dimensions had a comprehensive effect on frailty, further supporting that a comprehensive and individualized intervention is necessary to prevent frailty.