Optimized Hyperparameter Tuning of the Deep Neural Network Towards Heart Disease Classification on Feature Selected Using Particle Swarm Optimization

Abstract

               Heart disease is a complex disease in healthcare industry as it leads to large number of death around the world. Early detection and classification of the disease using computer vision technique increase the survival rate of the patients. Especially machine learning technique has been employed to electronic health records of the patients. Due to class imbalance, high dimensionality and data sparsity issue of the dataset, machine learning model finds it complex to detect and classify the disease with reduced accuracy. In order to manage those mentioned challenges, a new paradigm termed as deep learning architecture can be employed. In this article, a new optimized hyperparameter tuning of deep neural network architecture has modeled and implemented along the particle swarm optimization considered as feature selection technique is incorporated along the proposed architecture to generate the optimal feature towards classification and identifying the severity of the disease on the patients.  Further proposed method is constructed as oversampling technique to eliminate the overfitting issues and class imbalance issues on feature distributions. Hyperparameter tuning of deep neural network is to compute best hyperparameter for classification of the disease. Activation function of the model uses parameterized Rectified Linear Unit for forward propagation of the feature for classification with weight updating. Classification layer is regularized through batch normalization on setting the epoch value for training the dataset. Cross validation of the model on validation data is processed in loss layer using cross entropy function to achieve better accuracy in process of detection and classification.  Experimental analysis of the proposed model is evaluated using two bench mark dataset which represented as Cleveland dataset and ORBDA -electronic heath record. The results depicts that proposed model achieves high accuracy and less execution time on comparing against the state of art approaches. Model achieves 99.14 percent accuracy on the Cleveland dataset and 98.45 percent on the ORDBA dataset respectively.