Objective To construct and validate a prefrailty risk prediction model among community older people, so as to provide reference for early detection of prefrailty. Methods A total of 542 robust and prefrail community older people were screened to develop the prefrialty prediction model by using back propagation (BP) neural network machine learning.A second group of 205 robust and prefrail community older people were screened to validate the model performance using Received Operator Characteristics curve. Results The risk factors of prefrailty ranking in order of importance were age, hospitalization last year, fall last year, less exercise, multimorbidity, depression, cognitive function impairment, lower education, lower daily activity and polypharmacy.Compared with logistic regression model, BP neural network model had a better prediction performance, its AUC was 0.891, 95%CI(0.846-0.918), accuracy was 0.858, and specificity was 0.782. Conclusion BP neural network model has better prediction performance, and community workers could prevent the development of prefrailty in community older people through fall prevention, exercise intervention, chronic disease health education, depression and cognition intervention.