Transfer learning is the electrocardiogram reconstruction capstone

Electrocardiogram (ECG) reconstruction from reduced-lead configurations is essential for improving patient comfort and enabling wearable cardiac monitoring. Traditional reconstruction models (whether generic, population-specific, or patient-specific) require large datasets and extensive computational resources, limiting their practicality in real-world applications. This study investigates transfer learning as a strategy to overcome these limitations by enabling efficient personalisation of generic ECG reconstruction models. Three pipelines were evaluated: a linear regression model using leads I, II, V2, a wave-masked linear regression model using leads I, II, V3 (WMLR), and an ensemble of feed-forward neural networks. Generic models were trained on 10,000 normal ECG records from the CODE-15% dataset and fine-tuned using patient-specific data from the PTB-XL dataset. Performance was assessed using multiple metrics, including Pearson correlation, Dynamic Time Warping, Percentage Root-mean-square Difference, morphology similarity, and spectral similarity, across time intervals up to two years post-personalisation. Results show that transfer learning significantly improves reconstruction accuracy compared to generic models, with personalised models maintaining stable performance over extended periods. WMLR consistently outperformed other pipelines in correlation and morphological fidelity, demonstrating the continued relevance of linear approaches for resource-efficient deployment. While performance declined in cases where ECG morphology changed over time (like from normal to infarction), accuracy remained statistically acceptable, highlighting the robustness of transfer learning-based personalisation. These findings highlight the potential of transfer learning to enable scalable, long-term ECG monitoring systems that combine accuracy, adaptability, and computational efficiency.