Diagnosis of Diabetes-Mellitus in human given a chance to know a person that may be at risk of lengthened tricky situation. This work proposed to diagnosis diabetes mellitus based on a hybrid of carefully selected features method to reduce the features by combining Chi square test and Recursive feature elimination methods, so this was the first stage. Then the output of first stage is the input to the classification stage to obtain the true accuracy as the second stage. Logistic Regression(LR), K-NearestNeighbor(KNN), and Naïve-Bayes(NB) methods are used to classify nonappearance or appearance of diabetes mellitus disease. So the contribution of this research was to determine the optimal number of extracted and approved characteristics for rapid accurate diagnosis of disease by using two methods of information reduction and to benefit from the data collected in real in addition to standard data.This work has deal with PIDD from UCI and another dataset that collected from some patients. All the results is evaluated using accuracy(ACC), precision(PR), recall(RE), and f-score(FScore) measures. Algorithms LR, KNN, and NB achieved accuracy about (95% -98%) with combined feature selection method. So LR and NB achieved maximum accuracy(98%)with proposed method based on five and six features of 400 patient records, while KNN algorithm achieved accuracy (87%)with Chi-square test and Recursive feature elimination on diabetes-dataset.