Speckleplethysmographic (SPG) Estimation of Heart Rate Variability During an Orthostatic Challenge

Heart rate variability (HRV) provides insight into cardiovascular health and autonomic function. Electrocardiography (ECG) provides gold standard HRV measurements but is inconvenient for continuous acquisition when monitored from the extremities. Optical techniques such as photoplethysmography (PPG), often found in health and wellness trackers for heart rate measurements, have been used to estimate HRV peripherally but decline in accuracy during increased physical stress. Speckleplethysmography (SPG) is a recently introduced optical technique that provides benefits over PPG, such as increased signal amplitude and reduced susceptibility to temperature-induced vasoconstriction. In this research, we compare SPG and PPG to ECG for estimation of HRV during an orthostatic challenge performed by 17 subjects. We find that SPG estimations of HRV are highly correlated to ECG HRV for both time and frequency domain parameters and provide increased accuracy over PPG estimations of HRV. The results suggest SPG measurements are a viable alternative for HRV estimation when ECG measurements are impractical.

artifact and the noted time delay 11 . The differences are more easily noticeable in the frequency domain, especially in the high frequency band (0.15-0.4 Hz) 13 due to pulse transit time (PTT) variability and respiratory activity 12 . Novel phone-based reflectance PPG measurements provide improved accessibility but face the same inherent limitations as transmittance PPG as well as reduced signal quality 14,15 . Other state-of-the-art methods used for approximation of HRV lack feasibility outside of a lab-based setting 16 . Speckleplethysmography (SPG), an optical signal that measures changes in blood flow using laser speckle imaging 17 , provides an improved signal-to-noise ratio 18 and robustness in the presence of motion artifact and cold temperatures as compared to PPG 19 (Fig. 1). Similar to PPG, it can be measured from the finger and processed in real-time 19 . In addition, SPG peaks before PPG, which should improve accuracy and reduce the impact of vascular compliance on HRV estimation (Fig. 1). The components required for SPG acquisition, a budget camera and laser pointer, are relatively inexpensive 20 . To date, SPG has not been reported on in the literature as a measure of HRV. Given the aforementioned benefits of SPG, we determined the accuracy of SPG during an orthostatic challenge for estimations of HRV.

Results
Artifact correction. Although the subjects were asked to remain still, the impact of motion artifact on signal quality varied for each subject. Furthermore, the peak detection algorithm had higher accuracy for some subjects depending on noise inherent to the signals based on factors such as individual blood flow 21 and skin tone 22 . We noted the percentage of corrected artifacts in Kubios for each time series (Table 1; Fig. 2). As expected, the standing data had a larger percentage of artifacts. The seated PPG signals (black) had a larger number of corrected artifacts than the seated ECG (green) and SPG (blue) signals. An in vivo comparison of ECG (black), SPG (blue), and PPG (red) waveforms measured from a single subject. Because of the pulse transit time from the chest (ECG) to the finger (SPG/PPG), the ECG has its R peak before the SPG peak. SPG peaks before PPG.  www.nature.com/scientificreports www.nature.com/scientificreports/ time domain parameters. We independently compared the SPG and PPG estimates of HRV to the ECG HRV during both the sitting and standing periods of the orthostatic challenge. We plotted the standard deviation of normal to normal R-R intervals (SDNN; ms) from the 17 subjects, which reflects both the short-term and long-term cyclic components responsible for variability during the recording period 13 , given by the equation 23 :

ECG
In addition, we plotted the root mean square of the successive differences (RMSSD; ms), which reflects vagal tone 24 and short-term variability, from the 17 subjects, given by the equation 23 : We also generated the corresponding Bland-Altman plots (Fig. 3). The SPG results are highly correlated to the ECG results and fall within the predetermined acceptable limits of agreement for both the sitting (Fig. 3, column 1) and standing (Fig. 3, column 2) periods. The PPG results are less correlated to the ECG results (Fig. 3, columns 3 and 4) than the SPG results and some of the data points fall outside the acceptable limits of agreement. Interestingly, the PPG results improved when the subjects went from sitting to standing. for the 17 subjects during both sitting and standing conditions. SPG and PPG are compared to ECG on scatter plots with a line of best fit and the line of best fit equation, the Pearson's correlation coefficient, and significance in the upper left corner of the plot. Each Bland-Altman plot corresponds to the scatter plot directly above it, with the mean difference (black line), the 95% upper and lower limits of agreement (blue dashed lines), and the acceptable upper and lower limits of agreement (red dashed lines) also plotted. It should be noted that the Bland-Altman plots for SPG and PPG have different y-axis scales.
www.nature.com/scientificreports www.nature.com/scientificreports/ frequency domain parameters. In addition, we compared the low frequency (LF; ms 2 ) and high frequency (HF; ms 2 ) components of the three signals, which span the 0.04 Hz-0.15 Hz and 0.15 Hz-0.4 Hz bands, respectively 24 (Fig. 4).The low frequency band originates from long-term regulation mechanisms such as thermoregulation and hormonal mechanisms, while the high frequency band originates from vagal tone and relates to the respiratory cycle 24 . Once again, the SPG results are highly correlated to the ECG results during both standing and sitting. However, for the SPG versus ECG Bland-Altman LF plots from both the sitting and standing measurements, there is a single point that falls outside of the acceptable limits of agreement. All HF SPG measurements fall within the acceptable limits of agreement. Similar to the time domain measurements, the PPG measurements have a larger correlation to the ECG results when the subjects are standing. All of the PPG Bland-Altman plots have one or more points that fall outside the acceptable limits of agreement.
The last set of HRV parameters we compared combine the HF and LF components from above 25 . First, we examined the normalized HF (normalized units or n.u.), HF (ms 2 )/[HF (ms 2 ) + LF (ms 2 )]. Both the sitting and standing SPG measurements correlated well with the ECG measurements and better than the PPG measurements (Fig. 5, rows 1 and 2). The seated SPG measurements fell within the acceptable limits of agreement, but one point from the standing SPG measurements fell outside the acceptable limits of agreement. Next, we compared the LF/ HF ratio, which represents a mix of sympathetic and vagal activity 24 , between the three signals (Fig. 5, rows 3 and  4). Based on the Bland-Altman plots, SPG and PPG appear to underestimate the LF/HF ratio at higher ratios. SPG has a higher correlation with ECG as compared to PPG with ECG during both the sitting and standing conditions. for the 17 subjects during both sitting and standing conditions. SPG and PPG are compared to ECG on scatter plots with a line of best fit and the line of best fit equation, the Pearson's correlation coefficient, and significance in the upper left corner of the plot. Each Bland-Altman plot corresponds to the scatter plot directly above it, with the mean difference (black line), the 95% upper and lower limits of agreement (blue dashed lines), and the acceptable upper and lower limits of agreement (red dashed lines) also plotted. It should be noted that some of the Bland-Altman plots for SPG and PPG have different y-axis scales.

Discussion
HRV measurements provide rich information concerning autonomic regulatory capacity as an indicator of cardiovascular and neurological function. However, the gold standard technique for the determination of HRV parameters involves ECG, which is limited in situations where peripheral measurements are desirable due to convenience and motion artifact 8 . Substitute techniques that provide estimates of HRV parameters (i.e. PPG) address these limitations at the cost of accuracy 26 . Our results demonstrate that SPG estimates both time and frequency domain parameters of HRV with relatively high accuracy during both sitting and standing conditions, suggesting SPG could prove beneficial for remote measurements of ANS function. When examining the Bland-Altman plots, all SPG estimations remain within the acceptable limits of agreement for time domain measurements. Accuracy decreases for frequency domain measurements. The correlation coefficients decrease for the HF (n.u.) and LF/ HF ratio because of compounding errors. Our data suggest that SPG is more accurate than PPG across all HRV parameters assessed in this study when collected from the same device. As expected, a larger percentage of artifacts required correction during standing than sitting. One possible explanation for the improved PPG correlations during standing is that the signal quality improved such that the peaks were sharper, which allowed for improved peak detection when corrections were unnecessary 11 . The improved signal quality may be attributed to the increased blood volume in the finger due to gravity and location of the finger relative to the heart while standing 27 .
We acknowledge that the study does have some limitations. First, we did not control the room temperature, although it remained relatively constant for each subject measurement and did not shift more than 4 °C between subjects. Colder temperatures reduce the signal quality of PPG, which in turn makes accurate peak detection difficult 28 . Furthermore, we did not apply more complex signal processing methods for filtering of the signals prior www.nature.com/scientificreports www.nature.com/scientificreports/ to peak detection 29,30 . We deemed using SPG peaks to assist with PPG peak detection reasonable because both signals are collected by the same imaging device. With preliminary analysis, we observed that PPG peak detection results using the same filtering techniques applied to the SPG signal (described in the Methods section) were poor, possibly because the PPG signal has a smaller signal to noise ratio than SPG 18 . To aid in future motion artifact identification and correction, we suggest collecting accelerometer data during an orthostatic challenge 31 . The subjects involved in this study were aerobically trained athletes that may have a reduced minimum rise time 32 , which would improve HRV estimation accuracy; we acknowledge that the results of this study may not extend to the general population. PPG estimations of HRV become less accurate as subjects age due to increased arterial stiffness and more PTT variability, but we predict SPG would be more robust to these changes based on past results 19 .
SPG faces many of the same limitations as PPG because of the PTT separating thoracic electrical measurements at the heart from optical measurements at the fingertip. On the other hand, the reduced susceptibility to motion artifact and temperature during both sitting and standing conditions suggest SPG measurements are preferable to PPG measurements for estimating HRV. To the authors' knowledge, this is the first study to directly compare SPG and ECG as a substitute measurement for HRV, and the correlation coefficients obtained support the notion that SPG HRV estimations are preferable to PPG HRV estimations in settings when ECG HRV cannot be collected. SPG estimations of HRV can aid in the prevention of over-training by enabling remote and convenient monitoring of decreases in HRV. Furthermore, recent studies noticed a decrease in HRV after concussions 6,33,34 . SPG could provide a method for on-field monitoring of head impacts.

Materials and Methods
Subject recruitment. We recruited 17 healthy intercollegiate athletes (9 males, 23 ± 3.74 years; 8 females, 19.25 ± 1.28 years) who were undergoing preseason ECG measurements as part of a study designed to monitor athletes for head impact exposure. The subjects were instructed to avoid caffeine consumption for six hours prior to the measurement. All measurements were done in accordance with human subject protocols approved by the Institutional Review Board at University of California, Irvine (HS#2008-6307 and HS#2014-1338). Informed consent was obtained from all subjects. equipment. We utilized a commercial finger-clip blood-flow sensing device (Flowmet, Laser Associated Sciences (LAS), Inc., Irvine, CA) connected to a Microsoft Surface Pro 5 with LAS software for simultaneous SPG and PPG signal acquisition. A coherent light source (785 nm) transilluminated each subject's finger and a 752-pixel × 480-pixel CMOS array detected the transmitted light, similar to a pulse oximeter. The Flowmet acquired images at 250 Hz and the exposure time was adjusted to ensure adequate signal given different finger thicknesses and skin tones. For wireless ECG acquisition at 2000 Hz, we used a Nomadix Wireless Receiver with ECG Amplifier (BIOPAC Systems, Inc., Goleta, CA). We designed and built a circuit for optical triggering of the Flowmet and electrical triggering of the BIOPAC system to ensure temporal synchronization of the two monitoring devices.

Data collection.
We placed the Flowmet on the left index finger and the wireless ECG system in the lead II configuration on the chest of the subject. Next, we instructed the subject to remain seated still during the measurement with the room lights off and then triggered the data acquisition protocol for both devices. The room temperature ranged from 20 °C to 24 °C. We continuously collected SPG, PPG, and ECG data with the subject first seated for 5 minutes and then standing for 5 minutes (Fig. 6).

Data analysis.
We processed the raw ECG data using MATLAB software (R2018b, Mathworks, Inc., Natick, MA) for peak detection of the R wave and subsequent R-R interval calculation. The R-R intervals for seated and standing measurements were separated and saved as two text files for analysis in Kubios HRV Standard 3.1.0 (Kubios, Kuopio, Finland).
The Flowmet outputs raw data as average intensity, I, in camera counts from 0-255 and average speckle contrast squared, K 2 , from 0-1 for each image. We converted the average intensity to PPG according to the Beer-Lambert Law 35 : where PPG is measured in arbitrary units. Speckle contrast was converted to SPG, which correlates linearly with blood flow 36 , using the simplified speckle imaging equation 37 : where SPG is measured in arbitrary units and T is the exposure time of the Flowmet image detector.
To process the raw Flowmet data, we wrote MATLAB software for simple filtering and peak detection. We removed high frequency noise from the SPG and PPG signals using a 6 Hz low pass filter and the local DC components by subtracting the values from a 500-point (2 second) moving average 38 . Next, a third order, 11-point Savitzky-Golay filter was applied to the signals for smoothing without peak distortion 39 . We wrote peak detection software for the SPG signal to identify the first peak, located immediately after the peak of the first derivative, for consistency (Fig. 7). Since the PPG signal was generally noisier than the SPG signal, and both signals were acquired from the same device, the PPG peak was located by finding the first peak after the SPG peak (Fig. 7). We calculated the intervals between peaks and saved them as text files for further processing. For completeness, we compared results from the peak of the first derivative for both signals (Supplementary Figs S1-S4) and the foot of both signals ( Supplementary Figs S5-S8), defined as the trough immediately before the peak of the first derivative. The results for these two processing methods were less accurate for the SPG-based estimation of HRV.
We processed the text files of 5 minute intervals in Kubios with the default settings 23 . We manually corrected artifacts from the series of R-R intervals individually via the built-in thresholding function, without consulting the other available signals (e.g. the seated SPG signal was analyzed independent of the seated ECG signal for the same subject), to account for missed beats and poor peak detection. Artifact correction involved replacing the inaccurate R-R interval with a new interpolated interval based on the local surrounding intervals 23 . For calculation of the frequency domain results such as LF and HF, we applied the autoregressive approach, which has improved stability for shorter time series 40 . Statistical analysis. We plotted the Kubios HRV results from ECG, SPG, and PPG and then plotted the lines of best fit. We calculated the pairwise linear correlation coefficient for SPG versus ECG and PPG versus ECG. In addition, we generated corresponding Bland-Altman plots with 95% confidence limits of agreement 41 . We decided a priori that acceptable limits of agreement for HRV indices from SPG and PPG when compared to ECG would be within 20% variation of the mean ECG measurement 11 .

Data Availability
The datasets generated during and/or analyzed during the current study are available from the corresponding author upon reasonable request. Figure 6. Subjects remained seated for 5 minutes while a finger-clip device simultaneously collected SPG and PPG data and a wireless system collected ECG data. After a 30 second transition period for the subject to go from sitting to standing, 5 minutes of SPG, PPG, and ECG data were collected while the subject remained standing. Figure 7. SPG (blue) and PPG (red) data taken from a subject while seated. For consistency in SPG peak detection, the first peak after a major trough for each period was used for SPG N-N interval analysis. The first PPG peak after the nearest detected SPG peak was used for PPG N-N interval analysis.