Objective To build risk prediction models for cognitive frailty in community older adults based on different machine learning algorithms, to find out the best model, so as to provide reference for evaluating and preventing cognitive frailty of elderly people in community. Methods A total of 1 105 elderly people who underwent physical examination in three community health service centers in Suzhou were randomly divided into 773 cases in training set and 332 cases in verification set.Based on the bivariate logistic regression analysis result of the training set, six kinds of cognitive frailness risk prediction models were constructed using six machine learning algorithms (logistic, Bernoulli naive Bayes, random forest, extreme gradient lifting, k-nearest neighbor and support vector machine), and evaluated and compared in the verification set.Then a scoring form of cognitive frailty of elderly people in community was constructed based on the optimal algorithm and verified. Results Bivariate logistic regression analysis of training set identified 13 risk factors, the area under the receiver operating characteristic curve （AUC） of the six models was 0.8240.889, the sensitivity was 0.721-0.849, the specificity was 0.7000.837, and the Yoden index was 0.498-0.674.Random forest model was the best model.The score range of cognitive frailty scoring form for the elderly built based on this model was 0-180 points, the area under the ROC curve of internal and external verification was 0.858 and 0.831, and the optimal cut-off value was 75 points. Conclusion The prediction model based on random forest algorithm has the best prediction performance, while the model based on logistic regression has the worst prediction performance.The establishment of community older adults cognitive frailty scoring form based on random forest algorithm is helpful for community health workers to identify the elderly at a high risk of cognitive frailty and provide evidence for screening and intervention of cognitive frailty.