Session S63.5

Multi-Lead Wavelet-Based ECG Delineation on a Wearable Embedded Sensor Platform

N Boichat, V Barbero, N Khaled, F Rincon*, D Atienza

Complutense University of Madrid
Madrid, Spain

This work is dedicated to the sensible optimization and porting of a multi-lead (ML) wavelet-transform (WT)-based ECG wave delineator to a state-of-the-art commercial wearable embedded sensor platform with limited processing and storage resources. The original offline algorithm was recently proposed and validated in the literature, as an extension of an earlier well-established single-lead (SL) WT-based ECG delineator.
Several ML ECG delineation approaches, including SL selection according to various criteria and lead combination into a single root-mean-squared (RMS) curve, are carefully optimized for real-time operation on a state-of-the-art commercial wearable embedded sensor platform. Furthermore, these ML ECG delineation approaches are contrasted in terms of their delineation accuracy, complexity and memory usage, as well as suitability for ambulatory real-time operation. Finally, the robustness and stability of the considered ML ECG delineation approaches are benchmarked with respect to a previously validated single-lead implementation.
More specifically, 3 SL selection approaches are considered: a random selection, an optimal selection that consistently chooses for each delineation point the channel with less error (genie approach) and a sub-optimal training-based approach that identifies the single lead providing the minimum standard deviation of the delineation error, averaged over all delineation points. Finally, these approaches were compared with the ML combination of the leads into the RMS curve. We validated the performance of these approaches in real-time operation on the QT database (QTDB), which provides 105 2-lead records sampled at 250 Hz, with manual annotations of QRS complexes and P and T waves. The differences (mean+/-standard deviation in ms) between automatic real-time delineation on the embedded platform and the expert annotations for the faint P wave were as follows: the training-based lead selection achieved 13.1+/-13, slightly worse than the genie lead selection approach (7.7+/-11.1) and one sample better than the random lead selection (10.0+/-16.9). Finally, the RMS-based ML delineation was found to achieve 11.7+/-13.7, without requiring expert training for lead selection, yet critically depending on successful baseline wander correction of the multiple lead prior to combination.
In summary, this work investigates the optimization and porting to a commercial wearable sensor platform of two multi-lead techniques, in order to enhance the accuracy of the ECG signal delineation. Both techniques have been successfully validated in real-time on a standard manually-annotated database.

(Abstract Control Number: 67)