Deep learning enabled reconstruction of 12-lead ECG from three asynchronous smartwatch leads with 10% error and 0.85 Pearson correlation vs standard ECG.
Can deep learning models accurately reconstruct a standard 12-lead ECG from asynchronous single-lead smartwatch acquisitions in healthy subjects?
Deep learning algorithms can successfully reconstruct a standard 12-lead ECG from asynchronous single-lead smartwatch recordings taken at the wrist and torso with high fidelity.
Absolute Event Rate: 0% vs 0%
Abstract Background Smartwatch-based ECG acquisition is increasingly popular for health monitoring, but the typical single-lead configuration limits diagnostic accuracy. Recent advancement in bio-signal processing based on artificial intelligence methods, especially end-to-end encoder-decoder deep networks, have the potential to integrate the information conveyed by asynchronous single-lead acquisitions, obtained by placing smartwatch in different positions, to reconstruct 12-lead ECG. Purpose We demonstrate the feasibility of reconstructing 12-lead ECG with accuracy starting from multiple asynchronous single-lead acquisitions by using a smartwatch. We aim to use strategically end-to-end deep learning, split into two encoder-decoder networks, namely (1) recovering the synchronism among the leads, and (2) mapping the set of the three re-synchronized leads into the standard 12-lead ECG. Methods Eighteen healthy subjects (10 males, 8 females, age 24±3 years) were recruited (approved by Ethical Committee, Opinion n. 29/2021) for this study. The protocol involved the sequential acquisition of the ECG by smartwatch placed on the wrist (surrogate of lead I, duration 30 s, three repetitions) and in two additional locations on the torso (surrogate of lead II and lead V2). As a reference, each subject underwent concurrently the recording of a 12-lead ECG by standard equipment. The two encoder-decoder networks were trained on a random subset of subjects (80%) and validated on the remaining 20% of the participants. ECG signal quality was measured by basSQI and pSQI scores accounting for signal-to-noise ratio, and hosSQI score accounting for kewness and kurtosis were computed 1, 2. The level of fidelity of the 12-lead reconstruction was measured in terms of waveform percentage error and Pearson correlation coefficient. Results The signal quality of smartwatch ECG acquired on the thorax was obtained by integrating basSQI, pSQI and hosSQI scores: approximately 90%, out of 30-s long 162 chunks, were labeled as equivalent to high quality single-lead ECG signal. Deep-learning-based temporal alignment of asynchronous smartwatch ECG leads was effective as demonstrated by an improvement of time alignment by about 150%. Finally, the quality of the 12-lead reconstruction from three synchronized smartwatch ECG leads show a percentage error and Pearson correlation coefficient of 10% and 0.85, with respect to standard 12-lead ECG. Conclusions Our findings demonstrate that Huawei smartwatch-acquired signals from the torso maintain high quality and, when properly synchronized, can be used to reconstruct standard 12-lead ECG. This approach could contribute to more accurate and accessible cardiac monitoring.
Pagotto et al. (Sat,) reported a other. Deep learning enabled reconstruction of 12-lead ECG from three asynchronous smartwatch leads with 10% error and 0.85 Pearson correlation vs standard ECG.