L2G-ECG: Learning to Generate Missing Leads in ECG Signals using Adversarial Autoencoder.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Electrocardiogram (ECG) is vital for diagnosing cardiovascular diseases.
• Standard 12-lead ECG monitoring is not always feasible for trauma or injured patients.
• Generating complete ECG data from limited leads is a significant clinical challenge.

Purpose of the Study:

• To develop an adversarial learning-based method for generating missing ECG leads from a reduced set of available leads.
• To enhance the realism and diagnostic utility of synthesized ECG signals.

Main Methods:

• Utilized a convolution neural network-based encoder-decoder architecture.
• Employed a Visual Turing Test discriminator to improve signal realism.
• Trained and validated the model on the Physionet Challenge 2021 dataset, assessing performance with 1-7 available leads.

Main Results:

• Demonstrated successful 12-lead ECG reconstruction from as few as 5 leads (2 limb, 3 chest).
• Achieved significant improvements in average $AUP_{S}$ (0.81 to 0.98) and reductions in $AUP_{M}$ (30.96x10^-5 to 1.18x10^-5) as available leads increased.
• Quantitative analysis showed low Mean Square Error (MSE), high Structural Similarity Measure (SSIM), minimal information loss, and strong Lead-II feature generation.

Conclusions:

• The proposed adversarial learning method effectively generates realistic and diagnostically relevant 12-lead ECG signals from limited data.
• This approach offers a viable solution for ECG monitoring in patients where standard electrode placement is challenging.
• The model's performance highlights its potential for real-world clinical applicability in cardiovascular diagnostics.