Condition-fusion-and-hiding denoising diffusion model for electrocardiogram lead reconstruction.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiovascular Diagnostics

Background:

• Portable single-lead electrocardiogram (ECG) devices offer convenience but lack the diagnostic depth of 12-lead ECGs.
• Existing methods for reconstructing 12-lead ECGs from limited leads often neglect crucial diagnostic annotation information.

Purpose of the Study:

• To develop an advanced generative model that leverages annotated text and diagnostic conclusions to learn electrocardiogram (ECG) pathology.
• To reconstruct high-fidelity 12-lead ECGs from single-lead inputs, focusing on pathological waveforms and rhythms.

Main Methods:

• A novel "condition-fusion-and-hiding denoising diffusion probabilistic model" was developed.
• A two-stage training strategy was employed: condition-fusion for pathological learning using text and ECG signals, followed by condition-hiding for representation learning without text.
• The model reconstructs fixed-length, 10-second 12-lead ECGs from single-lead data.

Main Results:

• The proposed model demonstrated superior performance in reconstructing 12-lead ECG signals compared to state-of-the-art methods.
• Experimental results indicated enhanced reconstruction accuracy and improved classification consistency.
• The model effectively learned and emphasized pathological waveforms and rhythms during reconstruction.

Conclusions:

• The condition-fusion-and-hiding diffusion model successfully integrates diagnostic annotations for improved ECG reconstruction.
• This approach advances the capability of using limited-lead ECG data for comprehensive cardiovascular diagnosis.
• The method holds promise for enhancing the diagnostic utility of portable ECG devices.