Information Retrieval from Photoplethysmographic Sensors: A Comprehensive Comparison of Practical Interpolation and Breath-Extraction Techniques at Different Sampling Rates

Abstract

The increasingly widespread diffusion of wearable devices makes possible the continuous monitoring of vital signs, such as heart rate (HR), heart rate variability (HRV), and breath signal. However, these devices usually do not record the “gold-standard” signals, namely the electrocardiography (ECG) and respiratory activity, but a single photoplethysmographic (PPG) signal, which can be exploited to estimate HR and respiratory activity. In addition, these devices employ low sampling rates to limit power consumption. Hence, proper methods should be adopted to compensate for the resulting increased discretization error, while diverse breath-extraction algorithms may be differently sensitive to PPG sampling rate. Here, we assessed the efficacy of parabola interpolation, cubic-spline, and linear regression methods to improve the accuracy of the inter-beat intervals (IBIs) extracted from PPG sampled at decreasing rates from 64 to 8 Hz. PPG-derived IBIs and HRV indices were compared with those extracted from a standard ECG. In addition, breath signals extracted from PPG using three different techniques were compared with the gold-standard signal from a thoracic belt. Signals were recorded from eight healthy volunteers during an experimental protocol comprising sitting and standing postures and a controlled respiration task. Parabola and cubic-spline interpolation significantly increased IBIs accuracy at 32, 16, and 8 Hz sampling rates. Concerning breath signal extraction, the method holding higher accuracy was based on PPG bandpass filtering. Our results support the efficacy of parabola and spline interpolations to improve the accuracy of the IBIs obtained from low-sampling rate PPG signals, and also indicate a robust method for breath signal extraction.

Cardiorespiratory Assessments in Panic Disorder Facilitated by Wearable Devices: A Systematic Review and Brief Comparison of the Wearable Zephyr BioPatch with the Quark-b2 Stationary Testing System

“…Finally, Tsai et al [ 33 ] ( Table 1 ) used the wrist-worn smartwatch, Garmin Vivosmart 4 (Garmin International, Inc., Olathe, KS, USA), to continuously collect physiological data, including HR, activity levels, and the duration of different sleep stages, during a 1-year real-life study on 59 patients with a primary diagnosis of PD. The Garmin Vivosmart 4 provides continuous monitoring of HR through photoplethysmography (PPG), a non-invasive optical measurement method that uses a light source and a photodetector on the surface of the skin to measure volumetric variations in blood circulation [ 40 , 41 ]. The device also captures the steps taken, distance traveled, floors climbed, duration of wakefulness, and different sleep stages of its wearers.…”

Wrist-Worn Sensor Validation for Heart Rate Variability and Electrodermal Activity Detection in a Stressful Driving Environment

“…Thus, the longer these sequences, the higher the quality of the reconstructed E4 IBI series. Moreover, provided that the detection of the diastolic BVP foot points was performed on a conditioned and reconstructed version of the BVP signal, we are deeply aware that the BVP reconstruction algorithm we implemented undoubtedly induced interval approximation errors [43]. Despite these concerns, our primal purpose for this study was to try to use 100% of the recorded signal and to not exclude any portion of the raw BVP signals.…”

“…Also, SDNN reported moderate to high correlation values for each condition. However, according to the Bland-Altman analysis for SDNN (i.e., see Supplementary Material [42]), a non-negligible and statistically meaningful positive offset was introduced by the E4 device, both in Low-Risk and High-Risk Driving, which was likely due to the intrinsically higher signal-to-noise ratio of a wearable device [15], especially in dynamic scenarios, and also to the randomness increase caused by the peak approximation error [43]. In addition, some dependence on the experimental condition for the absolute error was found, mainly among those conditions where the motion frequency was considerably different, like between Baseline and Scream, Low-Risk and High-Risk Driving (i.e., the absolute error is lower in Baseline, as expected), or between Video Clip and Scream.…”

“…Although compared to other studies with similar experimental conditions [19], the mean artifact rates for E4 were significantly lower, the moderate to low correlation values of the E4-derived frequency domain parameters were mainly affected by the artifact correction procedure. Indeed, removing artifactual IBIs and then re-interpolating the missing values could have induced mismatches between the frequency domain outputs of the two measurement systems [43]. On that basis, experimental conditions with non-negligible movements (i.e., Low-and High-Risk Driving) were affected more than others.…”