Performance Analysis of Deep Learning Optimizers for Arrhythmia Classification using PTB-XL ECG Dataset: Emphasis on Adam Optimizer

Abstract

This study assesses the performance of several deep learning optimizers for arrhythmia classification via the PTB-XL ECG dataset. Deep learning (DL) based approaches, such as convolutional neural network (CNN) and recurrent neural network (RNN) have demonstrated promising results in learning discriminative feature representations of ECGs for automatic cardiac diagnosis. A CNN-LSTM based model was trained through six optimizers namely SGD, Momentum, Adagrad, Adadelta, RMSprop and Adam. The PTB-XL dataset with more than 20,000 12-lead ECGs was utilized for the classification performance comparison. Interest centred toward the Adam performance, which implies the adaptive moment estimated gradients and sets different learning rates for each learning rate. The Adam optimizer outperformed all other tested optimizers with 98.26% accuracy, 96.15% sensitivity, 97.43% specificity, 96.78% precision, and 96.46% F1-score. In comparison to other optimizers, they obtained low performance measures and reached convergence slowly. These results reveal the advantage of Adam in terms of training stability and predictive confidence for ECG-based arrhythmia classification. This research is one of the very few to systematically analyse various optimizers on PTB-XL dataset with hybrid architecture (CNN and LSTM). The experimental results confirm the superiority of Adam in ECG signal classification and provide a strong baseline for more effective deep learning model used in (cardiac) arrhythmia detection and clinical deep learning systems.