A Multi-Objective Bat Optimization Enhanced Convolutional Neural Network for Diabetic Prediction with Synthetic Data

Abstract

Diabetes mellitus (DM) is a serious health concern world wide as the population with DM raises considerably over years. This brings a necessity to predict DM very earlier with utmost accuracy. This research work provides a framework which combines Multi Objective Bat Algorithm (MOBA) which enhances the capability of Convolutional Neural Network (CNN). This MOBA-CNN is used to predict the DM earlier. Synthetic data which mimics the real world data is used for training the model. The MOBA is a kind of swarm intelligence which optimizes the feature subset by simultaneously maximizing the classification accuracy and bringing down the number of selected features. This enhances the predictive performance and computational efficiency. This work uses 50,000 synthetic data samples which includes several information like height, weight and other common diabetic risk factors like family history, diet etc. The generated samples are divided into training sets where MOBA optimizes the features and the optimized features are fed to the CNN. Experimental results portray that the optimized feature selection reduces the input dimensionality by over 30% while the critical information which is mandated for classification is preserved. The proposed framework achieved an average accuracy of 92.3% which surpasses the baseline models which operate without feature selection by 3-5%. In addition, the computational time is decreased by 18% due to the usage of feature optimization using MOBA. This speeds up the training process without compromising the prediction quality.