Deep Learning-Based Multi-Lead ECG Reconstruction from Lead I with Metadata Integration and Uncertainty Estimation.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiovascular Diagnostics

Background:

• Standard 12-lead electrocardiograms (ECGs) are crucial for cardiac diagnosis but require multiple electrodes.
• Single-lead ECG devices offer convenience but limited diagnostic information.
• Reconstructing 12-lead ECGs from single-lead data is an active area of research.

Purpose of the Study:

• To develop a novel dual-branch deep learning framework for high-fidelity 12-lead ECG reconstruction from single-lead inputs.
• To enhance reconstruction accuracy and clinical interpretability by integrating waveform data with metadata.
• To introduce predictive uncertainty estimation for improved reliability in ECG reconstruction.

Main Methods:

• A dual-branch deep learning architecture combining a CNN-BiLSTM for Lead I ECG signals and a fully connected network for clinical metadata.
• Utilized a dataset of 10,646 ECG records from a public repository.
• Applied Monte Carlo dropout during inference for predictive uncertainty estimation.

Main Results:

• The proposed framework, incorporating metadata, significantly outperformed the U-Net model in ECG reconstruction.
• Metadata integration improved reconstruction fidelity, particularly in QRS complexes and T-wave segments.
• Predictive uncertainty demonstrated a positive correlation with reconstruction errors, highlighting areas of reduced reliability.

Conclusions:

• Combining single-lead ECG waveform data with clinical metadata and uncertainty quantification is a promising approach for developing trustworthy wearable ECG systems.
• This study presents the first framework for ECG reconstruction that incorporates predictive uncertainty.
• The findings suggest enhanced potential for accurate and reliable remote cardiac monitoring.