Automated Arrhythmia Detection from ECG Signals: AComparative Study of Bidirectional Long Short-Term MemoryNetworks and Residual Convolutional Neural Networks

Technical Summary: Automated Arrhythmia Detection from ECG Signals

Problem Statement
Cardiovascular diseases remain the leading cause of global mortality, making the early and accurate detection of cardiac arrhythmias critical. While the electrocardiogram (ECG) is a standard diagnostic tool, traditional analysis relies on time-consuming, subjective visual interpretation by experts, which is prone to inter-observer variability. Although deep learning has revolutionized automated ECG analysis, there is a lack of direct, controlled comparisons between leading architectures—specifically Bidirectional Long Short-Term Memory (BiLSTM) networks with attention mechanisms and Residual Convolutional Neural Networks (ResNets)—on the same multi-class classification task. Existing studies often compare models across different datasets or with varying preprocessing steps, hindering fair evaluation and the establishment of robust baselines for clinical application.

Methodology
This study conducted a rigorous comparative evaluation of two deep learning architectures using a publicly available dataset of 109,446 single-lead ECG heartbeats (87,554 training, 21,892 test). The data was segmented into five classes: Normal (N), Atrial Premature Beat (APB), Premature Ventricular Contraction (PVC), Fusion of Ventricular and Normal Beat (F), and Unclassifiable Beat (Q).

• Preprocessing: Signals were normalized to a [0,1] range using min-max scaling. To address class imbalance, particularly for minority classes like Fusion beats, data augmentation was applied using Gaussian noise (mean=0, variance=0.01) and time-warping. The dataset was stratified into training (70%), validation (15%), and test (15%) sets.
• BiLSTM-Attention Model: This architecture utilized an input layer for 187-dimensional heartbeat vectors, followed by a single BiLSTM layer (64 units) processing sequences in both forward and backward directions. An attention mechanism was applied to weight the importance of different time steps, followed by a dense layer (64 units, ReLU) and a softmax output layer.
• ResNet Model: A 1D Residual Convolutional Neural Network was designed with an initial convolutional layer, followed by five residual blocks. Each block contained two 1D convolutional layers (32 filters, kernel size 5), batch normalization, and ReLU activation, connected via skip connections to mitigate vanishing gradients. Max-pooling layers reduced dimensionality after each block, followed by two dense layers and a softmax classifier.
• Training Protocol: Both models were trained under identical conditions using the Adam optimizer (learning rate 0.001) and categorical cross-entropy loss. The BiLSTM was trained for up to 50 epochs with early stopping (patience 10), while the ResNet was trained for 50 epochs.

Key Contributions

1. Systematic Benchmarking: The study provides a controlled, head-to-head comparison of BiLSTM-Attention and 1D ResNet architectures for multi-class arrhythmia detection, eliminating confounding variables such as dataset differences.
2. Performance Validation: It establishes ResNet as a superior baseline for beat-level classification, demonstrating its ability to extract complex morphological features more effectively than sequence-based approaches for this specific task.
3. Detailed Per-Class Analysis: The research offers granular performance metrics, highlighting specific strengths and weaknesses of each model across different arrhythmia types, particularly in distinguishing overlapping categories.

Results
The ResNet architecture significantly outperformed the BiLSTM-Attention model across all primary metrics:

• Overall Accuracy: ResNet achieved 98.0% validation accuracy compared to 95.7% for BiLSTM.
• F1-Score: ResNet attained a macro-averaged F1-score of 0.94, surpassing the BiLSTM's 0.88.
• Per-Class Performance: ResNet demonstrated near-perfect performance on Normal (N) and Unclassifiable (Q) classes (F1-scores of 0.99). Crucially, ResNet showed a marked advantage in the challenging "Fusion" (F) class, achieving a recall of 0.48 and an F1-score of 0.65, compared to BiLSTM's recall of 0.36 and F1-score of 0.38.
• Statistical Significance: The 0.06 difference in macro-averaged F1-scores indicates a substantial improvement, attributed to ResNet's capacity to learn highly discriminative morphological features through deep residual blocks.

Significance and Claims
The paper claims that these findings establish the 1D Residual Convolutional Neural Network as a more robust, stable, and reliable baseline for automated arrhythmia detection compared to BiLSTM-Attention models. The authors argue that while BiLSTMs excel at capturing temporal dependencies, the morphological complexity of ECG signals is better served by the deep feature extraction capabilities of ResNets.

The study concludes that ResNet is a powerful candidate for integration into next-generation healthcare technologies. However, the authors maintain a modest stance regarding immediate clinical deployment, acknowledging limitations such as the single-source nature of the dataset and the lack of inherent explainability in the models. They suggest that future work should focus on validating these models on multi-center datasets, exploring hybrid architectures (e.g., ResNet-BiLSTM), and developing explainable AI techniques to facilitate clinical adoption.