DR-net-Mamba: Selective State-Space Modeling for Long-Range ECG Time-Series Denoising

Problem

Non-stationary noise in ECG recordings significantly degrades diagnostic reliability, necessitating improved denoising techniques. This work addresses this gap by proposing a new model specifically designed for long-range ECG time-series denoising. The paper is a preprint and has not undergone peer review.

Method

The authors introduce a Mamba-augmented model that incorporates selective state-space blocks. This architecture is designed to combine local feature extraction with long-range temporal modeling, achieving linear complexity. The model's performance is evaluated using several metrics, including reconstruction fidelity, noise robustness, recording-length scaling, and downstream diagnostic classification. The training utilizes both synthetic and real datasets encompassing over 40 pathology classes, ensuring a comprehensive evaluation of the model's capabilities.

Results

The proposed model demonstrates superior performance across multiple metrics compared to existing denoising methods:

• SNR: Achieves the highest Signal-to-Noise Ratio (SNR) among evaluated denoisers.
• RMSE: Reports the lowest Root Mean Square Error (RMSE) compared to other denoising techniques.
• Macro AUROC: Attains the best macro Area Under the Receiver Operating Characteristic (AUROC) score among all denoisers, outperforming convolutional base models.
• Binary Cross-Entropy: Shows improvement over Inception1D when evaluated on noisy input.
• Brier Score: Also improves upon Inception1D in terms of Brier Score against noisy input.
• Calibration Metrics: The lead-specific Mamba variant uniquely enhances calibration metrics over the noisy baseline across both classifiers used in the study.

Limitations

The authors note that calibration performance is dependent on the classifier used, indicating that the denoising capabilities do not consistently outperform noisy input when evaluated with the ResNet1D-Wang model. This suggests potential limitations in the model's generalizability across different architectures.

Why it matters

The implications of this work are significant for the field of ECG analysis, as improved denoising techniques can enhance the reliability of diagnostic processes. By addressing the challenges posed by non-stationary noise, this model could facilitate better clinical outcomes and pave the way for further advancements in time-series analysis within medical applications.