A Transfer Learning Model -Based Patient-Specific ECG Lead Synthesis Algorithm

The 12-lead electrocardiogram (ECG) is a fundamental diagnostic tool in clinical cardiology, yet often, attaching the necessary electrodes can be cumbersome, particularly with wearable devices that are optimized for at-home monitoring and record a reducedlead set.Machine learning-based approaches have been employed to synthesize the remaining leads required for the 12-lead ECG, yet the generalizable nature of these models may not capture the distinctive morphology of a given patient's ECG.Herein, this study aims to incorporate a transfer learning-based personalized model to improve the accuracy of ECG lead synthesis on a perpatient basis.A generalized deep learning-based synthesis framework is initially developed, processing inputs from four measured ECG leads (I, II, V2, V4) to generate outputs for four synthesized leads (V1, V3, V5, V6).Trained instances of this model are subsequently fine-tuned on five full 12-lead recordings per patient, a notable reduction in the number of ECGs needed for personalization compared to similar work.Significant improvements in the personalized models over the generalized models were observed in patients with both normal and abnormal morphologies, indicating promising implications for the clinical application of personalized reduced-lead ECG lead synthesis.