Predicting diseases early plays an important role in improving healthcare and helping us prevent harmful health conditions earlier. Diabetes is the most contagious disease in the world, and serious complications include diabetic retinopathy, kidney problems, and heartburn. Machine learning (ML) models are used to collect relevant data from datasets. Data discovery relied heavily on outlier detection (OD). Identifying anomalous data within a given dataset is the main goal of outlier detection. We used optimal stochastic gradient descent (SGD) and multilayer perceptron (MLP) with outlier detection to create a model for predicting diabetes. This work goes into more detail about the proposed model. This study looks at the three attribute selection methods—backward elimination (BE), forward selection (FS), and stepwise regression (SR)—for choosing the features from preprocessed data. As a technique for identifying outliers, a radial basis function (RBF) system is used to extract misclassified samples. For high-efficient classification of data, the ML technique called MLP with maximum SGD was described. The result shows that using attribute selection and radial basis function improves the efficiency of diagnosing diabetes in terms of accuracy, sensitivity, and specificity.