Electrocardiogram lead conversion from single-lead blindly-segmented signals.

Area of Science:

• Biomedical Engineering
• Medical Informatics
• Artificial Intelligence in Healthcare

Background:

• The standard twelve-lead electrocardiogram (ECG) is crucial for cardiac diagnosis but requires numerous uncomfortable electrodes.
• Existing interlead ECG conversion methods often need multiple reference leads or precise temporal signal alignment.
• This limits the application of ECG in wearable devices and less obtrusive clinical settings.

Purpose of the Study:

• To investigate the feasibility of reconstructing all twelve ECG leads from a single reference lead without temporal alignment.
• To develop and evaluate a deep learning methodology for single-lead ECG signal conversion.
• To compare the performance of a U-Net architecture against convolutional autoencoders and label refinement networks.

Main Methods:

• A deep learning encoder-decoder U-Net architecture was employed for blind, single-lead ECG signal conversion.
• The methodology was tested using single shared encoders versus multiple individual encoders for each lead.
• Performance was evaluated against adaptations of convolutional autoencoders and label refinement networks.

Main Results:

• The proposed U-Net methodology achieved state-of-the-art performance in reconstructing multiple target ECG leads, even under challenging conditions.
• Leads I and II proved particularly effective as reference signals for converting specific lead sets.
• The method demonstrated promising cross-database performance and was robust to the presence of medical conditions.

Conclusions:

• Single-lead, blindly-segmented ECG conversion is feasible using deep learning.
• The developed methodology shows potential for improving cardiac health monitoring in wearable devices and clinical settings.
• Further research is needed to enhance robustness across diverse acquisition setups for broader applicability.