Arrhythmia Identification and Classification using Runge Kutta Optimizer-Based Hyperparameter Optimization for Long Short Term Memory

Abstract

Arrhythmia denotes to the abnormalities in the rhythm of the heartbeat experienced by individuals. The arrhythmia potentially causes fatal difficulties that lead to the sudden risk to life. Therefore, arrhythmia identification and classification is an important task for cardiac diagnosis. However, inappropriate hyperparameters initialized in the deep learning classifier fail to achieve full potential during the diagnosis. Therefore, this paper proposes an effective hyperparameter optimization using Runge Kutta (RUN) optimizer for a deep learning classifier i.e., Long Short Term Memory (LSTM). The important motivation of this research is to optimize the hyperparameters of LSTM for an effective arrhythmia identification and classification. First, the Electrocardiogram (ECG) signals of MIT-BIH arrhythmia dataset are transformed and decomposed using Multiscale Local Polynomial Transform (MLPT) and Ensemble Empirical Mode Decomposition (EEMD). Different features are obtained based on Standard Deviation (SD), Zero Crossing Rate (ZCR), Mean Curve Length (MCL), Hjorth parameters, Mean Teager Energy (MTE), and Log Energy Entropy (LEE) and ResNet-18. Next, a feature selection based on Improved Firefly Optimization Algorithm (IFOA) is used for choosing the optimum feature set. Therefore, the LSTM with optimized hyperparameters that uses RUN is utilized to enhance the classification. The proposed LSTM-RUN is analyzed based on accuracy, sensitivity, specificity, precision, and error rate. The existing researches, namely 2D-Convolutional Neural Network (CNN)-LSTM, deep learning and fuzzy clustering, namely Fuzz-ClustNet, and Extreme Learning Machine (ELM)-CNN are utilized for comparing the LSTM-RUN. The accuracy of LSTM-RUN is 99.87% which is higher than that of the existing approaches, namely 2D-CNN-LSTM, Fuzz-ClustNet, and ELM–CNN.