Inverse Filtering for ECG Signal Reconstruction-A Deep Learning Approach

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Wearable devices are crucial for large-scale cardiac arrhythmia screening, like atrial fibrillation.
• Signal conditioning in these devices can distort electrocardiogram (ECG) morphology, hindering diagnosis.
• Developing methods to reconstruct clean ECG signals is vital for accurate diagnostics.

Purpose of the Study:

• To propose a deep learning-based inverse filtering approach to reconstruct clean single-lead ECG signals.
• To overcome signal distortions caused by electrode characteristics in wearable diagnostic devices.
• To enhance the diagnostic utility of single-lead ECG devices for cardiac conditions.

Main Methods:

• A U-Net autoencoder deep learning model was employed for inverse filtering.
• A synthetic dataset was generated using MIMIC-IV and an object-oriented model simulating device characteristics.
• The model was trained on paired noisy and clean ECG signals.

Main Results:

• The U-Net autoencoder demonstrated effective reconstruction of clean ECG signals.
• Key ECG features were preserved with a Mean Squared Error (MSE) of 0.052 on validation and 0.134 on the test set.
• Real-world data evaluation showed improved extraction of temporal ECG features (PQ interval, QRS duration, QTc time).

Conclusions:

• The deep learning inverse filtering approach successfully mitigates signal distortions in single-lead ECGs.
• This method enhances the diagnostic accuracy of wearable cardiac screening devices.
• Improved extraction of temporal ECG features supports better clinical decision-making for arrhythmias like atrial fibrillation.