A Paralleled Multi-Task Learning-Based Framework for Single-Lead ECG Fine-Grained Noise Localization, Denoising and...

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Wearable electrocardiogram (ECG) devices are crucial for personalized healthcare but susceptible to transient noise during daily activities.
• This noise complicates signal classification, denoising, and can reduce diagnostic accuracy.
• Existing methods struggle with diverse noise types and maintaining waveform fidelity.

Purpose of the Study:

• To develop an advanced ECG preprocessing framework using multi-task learning.
• To improve the accuracy of ECG signal quality assessment and denoising.
• To enable robust, real-time ECG monitoring in wearable devices.

Main Methods:

• A Transformer-based multi-task learning model was developed, incorporating a fine-grained noise localization task.
• The model was trained using weak supervision and pathological ECG data, optimizing with three task-specific loss functions.
• Intra-class awareness was integrated to handle varied noise within quality categories, enabling adaptive denoising.

Main Results:

• The framework achieved state-of-the-art performance in ECG denoising and quality assessment, with F1-scores up to 98.49% and classification accuracy over 95.68%.
• Significant signal-to-noise ratio (SNR) improvement from -1.95 ± 3.83 dB to 12.20 ± 2.51 dB was observed under severe noise, preserving waveform fidelity.
• The model demonstrated effective compression via pruning and quantization for edge computing deployment.

Conclusions:

• The proposed method offers an efficient and clinically relevant solution for large-scale, real-time ECG monitoring.
• It preserves diagnostically important ECG waveforms and provides interpretable noise localization.
• The framework enhances the reliability and applicability of wearable ECG devices in personalized healthcare.