Wearable devices are rapidly spreading thanks to multiple advantages. Their use is expanding in several fields, from medicine to personal assessment and sport applications. At present, more and more wearable devices acquire an electrocardiographic (ECG) signal (in correspondence to the wrist), providing potentially useful information from a diagnostic point of view, particularly in sport medicine and in rehabilitation fields. They are remarkably relevant, being perceived as a common watch and, hence, considered neither intrusive nor a cause of the so-called "white coat effect". Their validation and metrological characterization are fundamental; hence, this work aims at defining a validation protocol tested on a commercial smartwatch (Samsung Galaxy Watch3, Samsung Electronics Italia S.p.A., Milan, Italy) with respect to a gold standard device (Zephyr BioHarness 3.0, Zephyr Technology Corporation, Annapolis, MD, USA, accuracy of ±1 bpm), reporting results on 30 subjects. The metrological performance is provided, supporting final users to properly interpret the results. Moreover, machine learning and deep learning models are used to discriminate between resting and activity-related ECG signals. The results confirm the possibility of using heart rate data from wearable sensors for activity identification (best results obtained by Random Forest, with accuracy of 0.81, recall of 0.80, and precision of 0.81, even using ECG signals of limited duration, i.e., 30 s). Moreover, the effectiveness of the proposed validation protocol to evaluate measurement accuracy and precision in a wide measurement range is verified. A bias of -1 bpm and an experimental standard deviation of 11 bpm (corresponding to an experimental standard deviation of the mean of ≈0 bpm) were found for the Samsung Galaxy Watch3, indicating a good performance from a metrological point of view.
Keywords: deep learning; electrocardiographic signal; heart rate; machine learning; measurement accuracy; measurement uncertainty; validation protocol; wearable sensors.