Electrocardiogram (ECG)-based wearable heart rate (HR) monitors provide a low-cost and accessible alternative to conventional multi-lead ECG systems, but their use in the development of medical applications is limited by the lack of annotated datasets. This study investigates methods for transforming standard 12-lead ECG recordings into single-lead signals representative of HR monitors, enabling the use of existing clinical databases.
A dataset of 24 synchronised recordings of 12-lead and single-lead ECGs, collected during graded exercise tests, was used to evaluate three transformation methods: linear regression (LR), feedforward neural network (NN), and one-dimensional convolutional neural network (CNN). Performance was assessed using signal similarity metrics and R peak detection and RR interval accuracy.
The CNN achieved the best results, with a correlation of 0.88, QRS detection sensitivity of 1.0, and RR interval error below 2 ms. The LR model showed slightly lower but still acceptable performance, while the NN underperformed. Only the CNN preserved signal morphology during arrhythmic events.
These findings demonstrate that CNN-based transformation enables realistic synthesis of HR monitor-like ECG signals and supports the reuse of existing ECG databases for developing algorithms for wearable devices.