Development and Validation of a Clinical Support System for Predicting Incident Frailty among Community-Dwelling Older Adults: A Prospective Cohort Study
Qinqin Liu,Liming Yang,Zhuming Shi,Jiaqi Yu,Huaxin Si,Yaru Jin,Yanhui Bian,Yanyan Li,Lili Ji,Xiaoxia Qiao,Wenyu Wang,Ming Zhang,Cuili Wang
DOI: https://doi.org/10.2139/ssrn.4170632
2022-01-01
Abstract:Background: Frailty becomes a great challenge with population aging. Early identification of older adults who are at high risk of frailty would facilitate appropriate and proactive health services.Objective: To construct prediction models and develop the frailty prediction system for identifying older adults at high risk of incident frailty.Design: Prospective cohort study.Setting and Participants: We used 2-wave data from the China Health and Retirement Longitudinal Study in 2013 and 2015. The physical frailty phenotype scale was used to determine frailty. A total of 2802 robust older adults aged ≥60 years were randomly assigned to a derivation cohort or validation cohort at a ratio of 8:2.Methods: The Least absolute shrinkage and selection operator was used to find the best subset of variables among 37 readily available variables for frailty prediction. Logistic regression, Random Forest, Support Vector Machine and eXtreme Gradient Boosting (XGBoost) were used to develop models for predicting incident frailty, respectively. The areas under the receiver operating curve (AUC), Hosmer-Lemeshow test, calibration plot, Brier score and decision curve analysis were used to evaluate the discrimination, calibration, overall performance and clinical usefulness of the prediction models. The frailty predictive system was web-based and implemented using Python, Flask, NumPy, Sklearn, uWSGI, Nginx and Gunicorn.Results: The incidence of frailty defined as frailty or prefrailty was 34.2% in derivation cohort and 34.8% in validation cohort. Predictors in the final model were: age, material wealth, housing tenure, current residence location, medical insurance, pension insurance, hearing impairment, lung diseases, self-rated health, waist circumference, eating habit, afternoon napping, depression symptom and cognitive function. Among the four models, XGBoost model achieved the slightly better prediction performance in terms of discrimination, calibration, overall prediction performance and clinical usefulness in both cohorts. The predictive ability (AUC) of XGBoost model was 0.700 (95% CI: 0.677–0.722) in the derivation cohort, while it was 0.701 (95% CI: 0.655–0.746) in the validation cohort. In XGBoost model, age, waist circumference and cognitive function were the top 3 important predictors for predicting incident frailty. We embedded the XGBoost model into the server and this clinical support system is accessible at http://www.frailtyprediction.com.cn.Conclusions: A user-friendly web-based system was developed with good performance using 14 readily available variables to assist healthcare providers to identify older adults at high risk of incident frailty and prompt frailty management. Further research is needed to test the effectiveness of using the system in a clinical environment.