ECG-Mamba: Cardiac Abnormality Classification with Non-Uniform-Mix Augmentation on 12-Lead ECGs

Objective: The detection of heart abnormalities using electrocardiograms (ECG) is a critical task in medical diagnostics. A lot of literature has utilized ResNet and Transformer architectures to detect heart disease based on ECG signals. Recently, a new class of algorithms has emerged, challenging these established methods. A selective state space model (SSM) called Mamba has exhibited promising potential as an alternative to Transformers due to its efficient handling of longer sequences. In this context, we propose a Mamba-based model for detecting heart abnormalities, named ECG-Mamba. Recognizing that common data augmentation methods such as MixUp and CutMix do not perform well with Mamba on ECG data, we introduce a data augmentation technique called non-uniform-mix to enhance the model’s performance. Methods and procedures: ECG-Mamba is based on Vision Mamba (Vim), a variant of Mamba that utilizes a bidirectional SSM, enhancing its capability to process ECG data effectively. To address the sensitivity of the Mamba model to noise and the lack of suitable data augmentation techniques, we propose a data augmentation algorithm that conservatively introduces data augmentation by performing non-uniform operations on the dataset across different epochs. Specifically, we apply MixUp to a portion of the dataset in different epochs. Results: Experimental results indicate that ECG-Mamba outperforms the best algorithms in the PhysioNet/Computing in Cardiology (CinC) Challenges of 2020 and 2021 based on the AUPRC and AUROC, specifically with ECG-Mamba achieving an AUPRC score 16.6% higher than the best algorithm in the PhysioNet/CinC Challenge 2021 on 12-lead ECGs, reaching 0.61. Moreover, with the proposed data augmentation method Non-Uniform-Mix, ECG-Mamba’s AUPRC reached 0.6271, representing a 2.8% improvement. Conclusion: The ECG-Mamba model, based on the SSM, demonstrates potential in detecting cardiac abnormalities from ECG data. Although the model surpasses existing algorithms, it exhibits sensitivity to noise, requiring careful data augmentation. The proposed conservative data augmentation technique addresses this challenge and improves the model’s performance, suggesting a promising direction for future research in ECG analysis using SSMs. The implementation is publicly available at https://huggingface.co/poult/ECGMambaVersionOfJTEHM2020-2021_final. Clinical and Translational Impact Statement: ECG-Mamba enhances heart abnormality detection, enabling early diagnosis and personalised treatment in resource-limited and telemedicine settings. Using real-world data from the PhysioNet/CinC Challenges 2020 and 2021, it accurately models multiple concurrent cardiac conditions, reflecting complex clinical scenarios. Its conservative Non-Uniform-Mix augmentation mitigates noise sensitivity, improving accuracy and reliability for seamless integration into clinical workflows, thus supporting evidence-based practice and addressing healthcare disparities.