Development of a clinical support system for identifying social frailty.

Abstract

Recognizing frailty, also known as clinical geriatric syndrome in the elderly that is characterized by high vulnerability and low resilience, and its extensive influence in clinical practice is challenging. This study aims to develop a social frailty prediction system based on machine learning approaches in order to identify the social frailty status of the elders in order to advance appropriate social services provision. This cross-sectional study enrolled and collected information from 595 community-dwelling seniors aged 65+. Fourteen predictors established from questionnaires and electronic medical records were used to predict the social frailty of participants. Bagged classification and regression trees, model average neural network, random forest, C5.0, eXtreme gradient boosting, and stochastic gradient boosting were used to build the predictive model in use. Performance was compared using accuracy, kappa, area under receiver operating characteristic curve, sensitivity, and specificity. The frailty predictive system was web-based and built upon representational state transfer application program interfaces. C5.0 achieved the best overall performance than remaining learners, and was adopted as the base learner for the social frailty prediction system. In terms of the area under receiver operating characteristic curve (AUC), health literacy (AUC = 0.68) was found to be the most important variable for predicting one's social frailty, followed by comorbidity (AUC = 0.67), religious participation (AUC = 0.67), physical activity (AUC = 0.66), and geriatric depression score (AUC = 0.62). Results suggest that a combination of such data that is both available and unavailable from electronic medical records is predictive of the social frailty of an elderly population.