Robust multi-class Type 2 Diabetes mellitus prediction model from imbalanced biomedical data using adaptive and attention-based residual BiLSTM

This study aims to develop a multi-class prediction model to provide reliable diagnosis outcomes for patients with type 2 diabetes mellitus. Publicly available data sources are used to acquire the patient data needed for the prediction of multi-class type 2 diabetes mellitus. Data preprocessing techniques such as outlier detection and filling are initially applied to the data to remove missing or inconsistent values to enhance the quality of the dataset. To address issues with data imbalance, ASMOTE is applied to the quality-improved data. The creation of synthetic minority samples is done to balance the class distribution of the data. The parameters present in ASMOTE are optimized using Modified Tasmanian Devil Optimization (MTDO) to improve data balance efficiency. Further, the multi-class classification is carried out using the developed Adaptive and Attention-based Residual Bidirectional Long Short-Term Memory (AA-RBiLSTM). By using the same MTDO, the hyperparameters of the RBiLSTM network can be tuned to further enhance the accuracy of this model. The prediction efficiency of our model is evaluated against prior works by using various performance measures. Adaptive Synthetic Minority Over-Sampling Technique (ASMOTE) to solve the data imbalance issues. The parameters present in ASMOTE are optimized using Modified Tasmanian Devil Optimization (MTDO) to improve data balance efficiency. Further, the multi-class classification is carried out using the developed Adaptive and Attention-based Residual Bidirectional Long Short-Term Memory (AA-RBiLSTM). From the performance evaluation, the suggested AA-RBiLSTM model is enhanced by 10.3% over LSTM, 4.8% over GA-SVM, 10.3% over DNN, and 6.9% over RBiLSTM for accuracy. The detection model using the proposed heuristic-based-deep learning prediction model is effective and improves the detection performance.