The advancement of artificial intelligence helps to predict and analyze huge quantities of data based on the supplies. Diabetic Retinopathy (DR) prediction and classification is an important research topic because, affording to a recent survey, around 422 million individuals worldwide are affected and are losing their vision. A plethora of studies and techniques have been developed to predict and analyze DR. However, until the main research gap is filled through the development of better models and the addition of new features, the advancement of machine and deep learning in terms of predictions and classifications changes many things in society. But the main problem in machine and deep learning is that the data is centralized, so the prediction and classification are performed only based on the central dataset, and the changes in the different places are not considered for prediction and classification. In this effort, we proposed a typical to sort out the mentioned issues such as central data storage-based training and training imbalanced problems. The proposed model is used in ensemble, convolutional neural network (CNN), and federated learning for prediction and classification in a distributed way. In the proposed work, ensemble learning is used to forecast the DR, the CNN algorithm is used to categorize the stages of DR, and federated learning is used to share the model with privacy and predict the result. The FedCS and Federated average are used for cleint selection and model averaging. The proposed work is implemented in the MESSIDOR dataset and checked for accuracy in the different scenarios of client selection and aggregation. Different scenarios are controlled with the help of automated privacy and information protection controls.