Transfer learning is the electrocardiogram reconstruction capstone.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning in Healthcare

Background:

• Electrocardiogram (ECG) reconstruction from limited leads is crucial for patient comfort and wearable cardiac monitoring.
• Conventional reconstruction methods demand extensive data and computational power, hindering practical application.
• Transfer learning offers a promising solution for efficient personalization of generic ECG reconstruction models.

Purpose of the Study:

• To investigate transfer learning for personalized ECG reconstruction.
• To evaluate the efficiency and accuracy of personalized generic models compared to traditional methods.
• To assess the long-term performance stability of transfer learning-based ECG reconstruction.

Main Methods:

• Trained generic ECG reconstruction models on a large dataset (CODE-15%).
• Fine-tuned models using patient-specific data (PTB-XL) via transfer learning.
• Evaluated three pipelines: linear regression, wave-masked linear regression (WMLR), and feed-forward neural networks.
• Assessed performance using correlation, Dynamic Time Warping, and morphology/spectral similarity metrics over two years.

Main Results:

• Transfer learning significantly enhanced ECG reconstruction accuracy compared to generic models.
• Personalized models demonstrated stable performance over extended periods (up to two years).
• WMLR pipeline showed superior correlation and morphological fidelity, highlighting linear models' efficiency.
• Accuracy remained acceptable even with ECG morphology changes, proving robustness.

Conclusions:

• Transfer learning provides an effective strategy for personalized ECG reconstruction, overcoming data and computational limitations.
• Personalized models offer scalable, accurate, and adaptable long-term ECG monitoring solutions.
• The WMLR approach demonstrates the viability of resource-efficient linear methods in personalized ECG reconstruction.