Predicting Chronic Diseases (CD) is crucial to healthcare informatics. Humans today suffer from a variety of diseases due to their lifestyle choices and their treatment of the environment. In order to prevent the severity of these diseases, it is crucial to recognize and foresee them in their early stages. Feature selection will reduce the amount of time required for training the model and enhance the accuracy of disease prediction. Improving the accuracy of classification systems depends extensively on feature selection. This research presents new feature selection and classification methods for CD diagnosis and prognosis. To increase the accuracy of disease diagnosis, Learning Dwarf Mongoose Optimization (LDMO) based feature selection is implemented by removing features that are less helpful or irrelevant. To enhance the searching capabilities, the LDMO algorithm with Learning Strategy (LS) is added; the revised alpha serves as a partial direction for the algorithm's updating process. The best essential feature that identifies the most useful aspects for CD prediction is selected using the LDMO method. A meta-heuristic technique called LDMO mimics the dwarf mongoose's compensatory behavioral modifications to mimic its foraging behavior. One particular kind of autoencoder is the Contractive Autoencoder (CAE). The penalty, which promotes the learnt representations to contract around the training is usually the Frobenius norm of the encoder activations to the input. The results demonstrate the superiority of CAE when compared to other methods based on precision, recall, F1-score, and accuracy.