Evolutionary-Based Feature Selection and Deep Recurrent Neural Network for Healthcare in IoTs System

Abstract

This paper proposes an optimized Deep Recurrent Neural Networks (DRNNs) for health care in IoT systems by selecting effective features and effective weight initialization based on genetic algorithm, to overcome the obstacles to training deep recurrent neural networks with restricted training data in feature space with high-dimensionality, such as over-fitting and vanishing/exploding gradients. Genetic algorithm is adopted for feature selection then the initial weights of the feature extraction layers are utilized to train the DRNNs, with the weights from the first stage of the feature selection layer being fixed. To enhance the network structures and determine the learning parameters, 10-fold cross-validation is employed. Performance evaluation of the trained DRNNs is conducted using test datasets not seen during the cross-validation process. Experimental results have revealed the effectiveness and advantages of the suggested approach.