A NOVEL EARLY DIAGNOSTIC FRAMEWORK FOR MULTI-CLASS TYPE 2 DIABETES MELLITUS PREDICTION WITH CLASS IMBALANCE USING SERIAL CASCADED DEEP ENSEMBLE NETWORKS

Significance of the study: Type 2 diabetes mellitus is a chronic disease that commonly occurs among aged people. So, early as well as personalized risk validation is considered essential for type 2 diabetes mellitus. Classical type 2 diabetes mellitus faces problems as it lacks the use of long-term temporal data and the collection of correlations among the diabetes risk factors and also, it does n’t validate the entire data. These kinds of unbalanced data complications presented in the classical type 2 diabetic mellitus prediction models need to be resolved to provide better disease-predicted outcomes. Aim of the study: A new multi-class type 2 diabetes mellitus prediction framework is developed to predict the class of disease, particularly when dealing with an unbalanced dataset. Also, a deep ensemble network is implemented with the novel concept to address the class imbalance issue presented in the classical type 2 diabetic mellitus prediction models by producing synthetic samples of minority classes that enhance the prediction performance. Methodology: Initially, the required data are collected from the standard source, and then preprocessing techniques such as outlier detection and data filling techniques are applied to transform the data into a usable format with higher quality. Then, the preprocessed data are given to Adaptive DeepSMOTE (ADSMOTE) to solve the data imbalance issues. Moreover, the Modified Frilled Lizard Optimization Algorithm (MFLOA) is utilized, which optimally tunes the parameters of ADSMOTE for improving the data balancing performance. Different classes of diabetes diseases are classified using Serial Cascaded Deep Ensemble Networks (SCDENs). Deep learning networks such as Bidirectional Long Short Term Memory (Bi-LSTM), Variational Autoencoder (VAE), and Convolutional Neural Networks (CNNs) are used for building SCDEN to classify the type of diabetes diseases. The classification accuracy of the developed ensemble model is better when compared to the single classifiers. Conclusion: Experimental evaluations are finally conducted to ensure the effectiveness of our proposed multi-class type 2 diabetes mellitus prediction technique. As a result, the proposed system achieves better performance with a greater accuracy of 96% at the epoch count of 400 in the early diagnosis of mellitus diabetes.