[Paper Review] ECGMamba: Towards Efficient ECG Classification with BiSSM

Yupeng Qiang, Xunde Dong+4 more|arXiv (Cornell University)|Jun 14, 2024

ECG Monitoring and AnalysisMedicine3 citations

TL;DR

ECGMamba proposes a novel, efficient ECG classification model using a bidirectional state-space model (BiSSM) based on the Mamba architecture to overcome the quadratic computational complexity of Transformers. It achieves state-of-the-art performance on two public ECG datasets with superior inference efficiency, demonstrating a strong balance between accuracy and speed.

ABSTRACT

Electrocardiogram (ECG) signal analysis represents a pivotal technique in the diagnosis of cardiovascular diseases. Although transformer-based models have made significant progress in ECG classification, they exhibit inefficiencies in the inference phase. The issue is primarily attributable to the secondary computational complexity of Transformer's self-attention mechanism. particularly when processing lengthy sequences. To address this issue, we propose a novel model, ECGMamba, which employs a bidirectional state-space model (BiSSM) to enhance classification efficiency. ECGMamba is based on the innovative Mamba-based block, which incorporates a range of time series modeling techniques to enhance performance while maintaining the efficiency of inference. The experimental results on two publicly available ECG datasets demonstrate that ECGMamba effectively balances the effectiveness and efficiency of classification, achieving competitive performance. This study not only contributes to the body of knowledge in the field of ECG classification but also provides a new research path for efficient and accurate ECG signal analysis. This is of guiding significance for the development of diagnostic models for cardiovascular diseases.

Motivation & Objective

• To address the high inference cost of Transformer-based models in ECG classification due to quadratic self-attention complexity.
• To explore the potential of bidirectional state-space models (BiSSM) for efficient and effective ECG signal modeling.
• To develop a lightweight yet powerful architecture that maintains high performance on long ECG sequences without sacrificing inference speed.
• To provide a new, efficient deep learning framework for clinical ECG diagnosis with improved scalability and computational efficiency.

Proposed method

• ECGMamba employs a Mamba-based block as the core component, leveraging its selective scanning mechanism for efficient long-sequence modeling.
• The model integrates residual connections and layer normalization to improve training stability and gradient flow.
• A feed-forward network (FFN) is applied after each Mamba block to enhance nonlinear representation learning.
• An ECG encoder based on convolutional layers is used for initial feature extraction before feeding into the Mamba blocks.
• The architecture uses a bidirectional design (BiSSM) to capture both past and future context in ECG sequences.
• The model is trained end-to-end with cross-entropy loss and optimized using standard deep learning training protocols.

Experimental results

Research questions

• RQ1 Can a BiSSM-based architecture outperform Transformers in ECG classification while maintaining low inference cost?
• RQ2 How does the integration of Mamba blocks with residual connections and normalization affect ECG classification performance?
• RQ3 What is the optimal number of Mamba layers for balancing accuracy and efficiency in ECG classification?
• RQ4 How do individual components (ECG encoder, LN, FFN) contribute to the overall model performance?

Key findings

• ECGMamba achieved an AUC of 0.9526 and F1-score of 0.7909 on the CPSC2018 dataset, outperforming baseline models.
• On the PTB-XL dataset, ECGMamba reached an AUC of 0.9167 and F1-score of 0.7528, demonstrating strong generalization.
• The ablation study confirmed that the ECG encoder, layer normalization, and FFN each significantly improved model performance.
• Using 8 Mamba layers yielded the best performance, with AUC of 0.9539 on CPSC2018 and 0.8881 accuracy on PTB-XL.
• The model maintained high inference efficiency due to the linear complexity of the Mamba mechanism, unlike the quadratic complexity of self-attention.
• The results indicate that BiSSM-based models are a viable and efficient alternative to Transformers for ECG classification.