Smartphone-enabled pulse rate variability: An alternative methodology for the collection of heart rate variability in psychophysiological research

International Journal of Psychophysiology 89 (2013) 297–304

Contents lists available at ScienceDirect

International Journal of Psychophysiology journal homepage: www.elsevier.com/locate/ijpsycho

Smartphone-enabled pulse rate variability: An alternative methodology for the collection of heart rate variability in psychophysiological research☆ James A.J. Heathers ⁎ The University of Sydney, Sydney, Australia

a r t i c l e

i n f o

Article history: Received 1 February 2013 Received in revised form 7 May 2013 Accepted 27 May 2013 Available online 8 June 2013 Keywords: Heart rate variability Pulse transit time Photoplethysmography Smartphone

a b s t r a c t Heart rate variability (HRV) is widely used to assess autonomic nervous system (ANS) function. It is traditionally collected from a dedicated laboratory electrocardiograph (ECG). This presents a barrier to collecting the large samples necessary to maintain the statistical power of between-subject psychophysiological comparisons. An alternative to ECG involves an optical pulse sensor or photoplethysmograph run from a smartphone or similar portable device: smartphone pulse rate variability (SPRV). Experiment 1 determined the simultaneous accuracy between ECG and SPRV systems in n = 10 participants at rest. Raw SPRV values showed a consistent positive bias, which was successfully attenuated with correction. Experiment 2 tested an additional n = 10 participants at rest, during attentional load, and during mild stress (exercise). Accuracy was maintained, but slightly attenuated during exercise. The best correction method maintained an accuracy of +/−2% for low-frequency spectral power, and +/−5% for high-frequency spectral power over all points. Thus, the SPRV system records a pulse-to-pulse approximation of an ECG-derived heart rate series that is sufficiently accurate to perform time- and frequency-domain analysis of its variability, as well as accurately reflecting change in autonomic output provided by typical psychophysiological stimuli. This represents a novel method by which an accurate approximation of HRV may be collected for large-sample or naturalistic cardiac psychophysiological research. Crown Copyright © 2013 Published by Elsevier B.V. All rights reserved.

1. Introduction Heart rate variability (HRV), the quantification of beat-to-beat variability in the cardiac cycle over time, is one of the most commonly used measures in psychophysiological research, and is employed broadly as a determinant of the status of the autonomic nervous system (ANS). To date, it has proved to be an important index of individual differences associated with emotional regulation (Thayer et al., 2012), mood and affective disorders (Grippo and Johnson, 2002), personality (Ode et al., 2010), and other facets of individual differences. The conventional data collection method of a short-term heart rate (HR) series involves a dedicated electrocardiograph (ECG), run from a computer or microcontroller, which is attached to a single participant in a laboratory setting. HRV metrics are calculated from the time series of R-waves provided by the ECG, which are the signal antecedents of ventricular depolarisation over time. While such a collection procedure is most common in published research, it has two central and related limitations. Firstly, a dedicated ECG system requires data to be collected singly, and relatively little attention has been paid to the development

☆ Funding source: The author was supported by an Australian Postgraduate Award scholarship at the University of Sydney. ⁎ The University of Sydney, Building A18, Griffith-Taylor Building, The University of Sydney, 2006, Sydney, Australia. Tel.: +61 293513544. E-mail address: [email protected].

of a flexible system designed specifically to collect HRV on the scale required for individual differences research. Traditionally, individual differences research involves questionnaire-based data or psychometric tasks administered simultaneously in multiple participants, within large samples. Furthermore, many individual differences studies, particularly with special groups (e.g., children, the elderly), must occur outside conventional laboratory settings. Secondly, the reliable calculation of HRV faces a number of methodological and procedural barriers, such as the need for careful data handling procedures to prevent error (Berntson and Stowell, 1998), profound distortions introduced to HRV values by non-fasted participants (Lu et al., 1999; Routledge et al., 2002), the need for monitoring breathing rate to quantify the frequency band respiratory sinus arrhythmia (Bernardi et al., 2000), disagreement over the necessity of controlled breathing (Grossman and Taylor, 2007, but see Denver et al., 2007), and difficulty of cross-comparison between non-standardised analytical methods (Sandercock, 2007). A further challenge is the internal reliability of these methods, and several attempts have been made to characterise this for standard HRV metrics over the immediate or short-term (Pitzalis et al., 1996; Sandercock et al., 2005; Pinna et al., 2007). For example, Pinna et al. (2007) report that to detect a standardised difference between groups (i.e., where HRV metrics differ by 30% of their between-subject standard deviation), adequate power in a free-breathing sample is achieved between n = 30 (using logHF power according to the Blackman and Tukey (1959) method) to n = 77 (using a log-LF/HF ratio method). In analyses that involve

0167-8760/$ – see front matter. Crown Copyright © 2013 Published by Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.ijpsycho.2013.05.017

298

J.A.J. Heathers / International Journal of Psychophysiology 89 (2013) 297–304

between group comparisons, however, it is also necessary to consider the division of the available sample into subgroups: recent work (Simmons et al., 2011) suggests a sample size of 20 observations per cell. Finally, these metrics are typically compared to psychometric inventories, which typically display moderate reliability themselves (e.g. Deditius-Island and Caruso, 2002). These challenges in defining the a priori environment, from which we might determine a required sample size, are best addressed by increasing the cell size of comparisons to HRV metrics. This work addresses the problems above through the collection of HRV from a smartphone platform, that is, from a mobile telephone or similar system with the provision of an operating system that allows access to computing and data transfer functions. Pew Research reports around two-thirds of young adults (and half of American adults overall) own a smartphone, and that ownership increased by more than 10% from May, 2011 to September, 2012 (Pew Research, 2012). While smartphones have been previously utilised for telehealth applications (see Free et al., 2012a,b for review), the platform is also increasingly being used for health and medical hardware development.1 In this context, the large and growing uptake of the smartphone platform solves the problem of individual access to either a computer or similar dedicated microprocessor and provides a dedicated environment for hardware/software development. From the smartphone platform, an optical recording of the pulse wave typically referred to as photoplethysmography is taken. This technique is the most common alternative to the ECG measurement of cardiac cycles, and can be used to derive an approximation of beat-to-beat heart period and thus calculate HRV. First outlined by Hertzman (1938), this method relies on transforming the pulsatile waveform of microvascular blood flow from a peripheral site on the body (typically the finger or ear) into a series of pulse-to-pulse (PP) intervals. This occurs via a simple device consisting of a near-infrared photodiode and a receiver; the presence of the systolic beat produces a perturbation in the light's absorbance which is identified as a pulse beat. Several groups have reported the accuracy between simultaneously measured pulsatile and ECG sources to be sufficient to estimate HRV (Bolanos et al., 2006; Gil et al., 2010; Heathers et al., 2012; Lu et al., 2009). For instance, Gil et al. (2010) report a correlation at supine rest between both low and high-frequency power, as both raw power and normalised units — n.u., in the frequency domain at or above r = 0.996. Errors were commensurate to this (e.g. ECG HF n.u. 34.4% vs. Pulse HF n.u. 34.7%). Optical sensors are also inexpensive, portable, computationally efficient, completely non-invasive, reusable, and – most importantly – low powered. Consequently, optically-derived pulse rate is an ideal candidate for mobile/smartphone collection of an approximation to HRV metrics: Smartphone Pulse Rate Variability (SPRV).

1.1. The current studies The SPRV system is presented as a solution for the accurate mass provision of between-subjects HRV data within two experiments. Experiment 1 outlines the basic parameters of the SPRV system, the accuracy of the raw data collection and methods of improving the fidelity of the pulse signal. While some implementations of pulse-ECG comparisons are accurate, the accuracy of any individual comparison is dependent on its individual implementation. Thus, the utility of the pulse signal to approximate the electrocardiograph is not always supported. For example, Charlot et al. (2009) report only moderate accuracy at supine rest (ECG HF n.u. 28% vs. Pulse HF n.u. 31%), with accuracy degrading progressively under both postural stress and during exercise. Wong et al. (2012) report poor accuracy between ECG and pulse signals taken at multiple sites (e.g. ECG HF n.u. 33.9% vs. Pulse HF n.u. 40.9%). 1

See, for instance, the Masimo iSpO2 — http://ispo2.com/.

These inaccuracies are at least in part due to the fact that pulseto-pulse intervals only approximate rather than directly measure the cardiac cycle. That is, the pulse signal is dependent on the vascular environment that exists between the initiation of the pulse wave at the heart, and the measurement of the pulse wave at the peripheral site; the pulse transit time. This interval has a systematic, moderately reliable relationship with systolic blood pressure (Obrist et al., 1979; Poon and Zhang, 2006; Payne et al., 2006) to the extent that it has been used as a psychophysiological measurement in its own right. As the same autonomic and circulatory factors which affect pulse transit are also present in the heart signal, it is theoretically possible to correct a pulse signal in order that it more closely conforms to the underlying ECG signal. A previous attempt to compare a traditional optical pulse monitor at rest to an ECG is instructive (Giardino et al., 2002), providing basic comparisons, measures of proportional and overall bias, and the direction of this potential bias, in both raw heart series and common HRV calculations. Experiment 2 replicates the initial accuracy of these initial solutions with a separate sample at rest, and confirms this accuracy during psychophysiologically relevant active conditions (i.e. during high attentional load and during exercise stress). 2. Methods 2.1. Participants Ten (10) adults (range = 21–30; M = 25.5, SD = 3.5; 6 M, 4 F) participated as volunteers in Experiment 1, followed by a separate sample of ten (10) (range = 18–28; M = 23.3 SD = 2.9; 7 M, 3 F) in Experiment 2. All participants were volunteer undergraduate or postgraduate students from the University of Sydney, and reported normal/corrected-to-normal vision, no vascular or neurological illnesses, and no regular medication other than the contraceptive pill. All participants viewed, signed and returned a Participant Information Sheet approved by the local Institutional Review Board. 2.2. Measurements 2.2.1. ECG Ag/AgCl electrodes were attached in a modified Lead-II formation (right clavicle and left iliac crest) with a reference electrode on the left clavicle, and connected to a laboratory ECG sampling at 1000 Hz (PowerLab 8/30: ADInstruments, Sydney, AUS) 2.2.2. SPRV system The SPRV system consists of a sensor (iThlete Finger Sensor: HRV Fit Pty Ltd, Hampshire, UK), consisting of a matched IR LED and photodiode embedded in an FDA-approved compliant silicone finger clip, with the light source transmitted above the eponychium, through the finger, and received on the pad of the distal phalange. Participants chose their own finger site for comfort and ease of use. The signal from the photodiode is fed into an interface box, which is connected to the microphone input of an iOS-compliant smartphone (i.e. an iPhone, iPad or iPod Touch), digitised at 16-bit resolution, lowpass filtered at 5 Hz and re-sampled at 500 Hz to give a 2 ms time domain resolution using custom software. The system defines the pulse fiducial point as the largest negative dA/dt in the same manner as Heathers et al. (2012). An adaptive threshold (Zong et al., 2003) is used to gate and separate the desired peaks in the dA/dt waveform from other smaller peaks (for instance, at the dichrotic notch) to return a pulse-to-pulse (PP) series. An animation of a beating heart on the application screen displays these beats in real time, thus both the experimenter and naïve participants can immediately interpret the correct function of the device. The entire device weighs approximately 25 grams, and can be seen in Picture 1.

J.A.J. Heathers / International Journal of Psychophysiology 89 (2013) 297–304

2.2.3. Tasks In Experiment 2, tasks additional to the rest period were deployed: 2.2.3.1. Choice reaction time go/no-go task. Participants were required to assort coloured circles (green/red) by keypress, and inhibit responding to two similarly coloured squares presented in random order. The frequency of the go vs. no-go responses was 2:1 and changed every 20 presentations, with the total frequency approximately equal. The inter-trial interval was randomly 400, 600, 800, 1200, or 1600 ms, where the two shortest latencies were twice as common. Wrong answers were punished with a 500 ms burst of white noise delivered through directional speakers at approximately 85 dB. Participants were urged to perform trials as fast as possible while maintaining accuracy, thus the task required sustained attention to avoid errors. Pilot testing confirms this task, which requires sustained low-level attention and provokes errors due to the violation of expectancy with the variable trial latency, conforms to the literature on the psychophysiology of attention which shows focused, attentionally demanding or time-pressured tasks provoke a reliable decrease in LF HRV power (Mulder and Mulder, 1987; Meijman, 1997; Aasman et al., 1987; Mulder et al., 1993; Middleton et al., 1999; Nickel and Nachreiner, 2003) 2.2.3.2. Basic exercise task. Participants were required to pedal on an under-desk cycle ergometer purpose-built to allow participants to remain in precisely the same seated position during the task. Cadence was controlled by electronic metronome. During the task, participants began pedalling at 80BPM with no external resistance, and heart rate was calculated every 20 s. Cadence was increased by 4BM if the decrease in heart period was less than 100 ms, and the recording period (5 min) commenced when the heart period was at least 100 ms reduced from baseline. No participant required a cadence over 116BPM. Exercise provides an unambiguous increase in sympathetic output, immediately increasing HR, cardiac output, and reliably reduces LF and HF power (Sandercock and Brodie, 2006).

299

Experiment 2 — in the 30 conditions collected (n = 10, with recordings at rest, during attentional load and during exercise), errors in Rwave detection in the ECG trace were linearly hand-corrected for four participants in seven conditions. No trace required more than 1% of beats to be corrected. Frequency domain measures approximating the activity of respiratory sinus arrhythmia (log-adjusted high frequency, 0.15 Hz– 40 Hz; lnHF) and the outflow of the baroreflex (log-adjusted low frequency, 0.04 Hz–0.15 Hz; lnLF) were calculated by Fast Fourier Transform using Welch's Periodogram (window width 256 s, 50% overlap, interpolated at 4 Hz, natural log-transformed). HRV metrics were calculated at every point via Kubios (v2.0, Biosignal Analysis and Medical Imaging Group, University of Kuopio, Finland). Time domain measures (SDNN: standard deviation of normal-to-normal intervals; RMSSD root mean squared successive difference) were also calculated as previously defined (Task Force, 1996). Neither of these measures violated Shapiro–Wilk's test for normality, thus raw scores were used. To characterise the nature of the error in Experiment 1, Spearman's correlations were calculated between the residual error series and the SPRV series, as the ECG may assumed accurate, and these methods were compared according to Williams' (1959) modified Hotelling statistic. To account for any potential error in difference plots due to this method (Bland and Altman, 1995), Bland–Altman plots were first calculated, with bias/limits of agreement (LoA) reported (Bland and Altman, 1986). First-order regression coefficients were also calculated to quantify potential directionality in potential bias. Bland–Altman values, regression equations and all plots were generated in Prism v.5 (Graphpad Software, San Diego, CA). All other analysis and correction was performed in MATLAB r2010b (MathWorks, Natick, MA). For Experiment 2, expected task effects were analysed with directional (one-tailed) or non-directional (two-tailed) t-tests, dependent on prediction (Student's t, within-subjects). Other HRV calculations and analyses were performed as before. All values are given as mean +/− SEM.

2.3. Procedure 3. Results In all rest conditions, participants sat at rest for 10 min with both monitors attached, the final 5 min of which was recorded simultaneously by both monitors and defined as the period of interest. This time period is taken as standard in short-term HRV recordings, as it contains sufficient cycles to successfully resolve low-frequency power (0.04 Hz). The device was kept out of sight to avoid potential biofeedback from the display after the fidelity of the pulse signal was established. If obvious errors were displayed, the task was extended or repeated until the recording period was met. During recording, the angle of the upper arm was maintained static and at rest, as changes in arm height and elbow angle have been observed to modify the pulse transit time (Foo et al., 2005). Tasks were performed as per 2.2.3 in set order (rest, attention, exercise) to prevent exercise from modifying resting-state HRV, and participants rested for 5 min between tasks. 2.4. Analysis ECG RR series was obtained by the identification of the zero-points after a local maximum of the first derivative series via dedicated software (HRV Module, Labchart, ADI). SPRV PP series was obtained via application export, which appends the phone log as an attachment to an email in the mail client of the smartphone. Consequently, the successive difference of the time series of pulse waves returns the PP series. One data record from Experiment 1 was truncated by 22 s to remove a movement artefact from the anticipation of completion. All other beats were of sinus origin and no other correction was necessary. Data quality was lower in

3.1. Experiment 1: initial analysis Close agreement between series was found for all participants. Correlation coefficients (Pearson's r) between RR and PP series were high for all participants, 9 out of 10 series were r ≥ 0.99 (range: 0.988–0.999, median: 0.996). Fig. 1 includes a representative (i.e. the median) series and an exploded section representing the scale of the overall error alongside a residual error series (where RR values were treated as correct). Table 1 summarises the agreement of common HR and HRV metrics calculated between ECG and SPRV systems. All HRV values from SPRV are significantly greater than ECG (paired-samples t-tests; for all comparisons, t9 > 5, p b 0.001). Fig. 2 shows the difference plot for a representative PP series against the error to ECG, which displays a small overall bias and acceptable limits of agreement, but a clear bias directionality — error increased positively with longer heart period and vice versa. This was extremely consistent, as the first-order coefficient of linear regressions of the Bland–Altman plots were significant in 9 out of 10 participants. Nearly all correlations between the residual error series and all ECG/SPRV series were significant (smallest significant value Pearson's r = 0.115, p = .044) and were significantly higher with SPRV than ECG values (Williams' Hotelling t for dependent correlations, all t9 > 10, all p b 0.001) i.e. it may be assumed that smaller beats were significantly underestimated by the SPRV system, and longer beats overestimated. These results are summarised for individual participants in Table 2.

J.A.J. Heathers / International Journal of Psychophysiology 89 (2013) 297–304

Error (ms) Heart period (ms)

300

1000

ECG value

1200

Heart period (ms)

SPRV value

900 800

10 5 0 -5 -10

1000

50

55

60

65

800

0

100

200

300

Time (s) Fig. 1. The ECG vs. SPRV comparison for a representative subject over 5 min. The exploded section shows an expanded section of this comparison above the error between the two traces. ECG = electrocardiograph; SPRV = smartphone-mediated pulse rate variability.

Initial data in Experiment 1 clearly indicated that the raw SPRV signal overestimates the length of proportionally longer heart periods, and underestimates shorter periods. Therefore two methods of correction were attempted: a) if stationarity is assumed, then the mean and variance of the heart rate are stable over the measurement period, and the size and direction of the error are directly proportional to the size of the beat. To correct for this, the SPRV series was z-score transformed, the product of the z-series and the median standard deviation of the positive bias was taken, and the sum of the raw and z-corrected series was taken as the final value. As this represents a correction with the assumption of stationarity, the series is termed S-SPRV. b) the timing of the heart beat over time has frequently been characterised as a stochastic process (e.g. Goldberger et al., 2002), which implies the insufficiency of spectral analysis to fully characterise heart rate variability as a series of stable, period processes. If this is the case, the mean and variance against which any frequency is observed may fluctuate over time. To attempt to correct for this, the SPRV series was cubic spline interpolated (10 Hz), and a correction series was calculated where a Savitsky–Golay polynomial filter

Table 1 Basic comparison of ECG vs. SPRV values. Mean RR = average heart rate; SDNN = standard deviation of normal-to-normal intervals; RMSSD = root-mean squared of successive differences; ln LF = natural logarithm of low-frequency spectral power (0.04 to 0.15 Hz); ln HF = natural logarithm of low-frequency spectral power (0.04 to 0.15 Hz); Data is given as mean +/− SEM. HRV metric

Mean RR (ms) SDNN (ms) RMSSD (ms) Ln LF_WP (ln ms2) Ln HF_WP (ln ms2 )

n

10 10 10 10 10

ECG (X1)

847.5 55.8 36.7 7.24 5.86

(44.7) (7.3) (5.3) (0.34) (0.29)

SPRV (X2)

847.9 57.6 38.3 7.31 6.04

(44.8) (7.5) (5.3) (0.34) (0.26)

Residual error (X2 − X1)

Coefficient of variation

0.42 1.79 1.61 0.07 0.18

0.04% 2.59% 3.43% 0.74% 2.69%

(0.06) (0.35) (0.28) (0.01) (0.05)

was used to remove the underlying trends below approximately 0.06 Hz for LF (167-point width, 3rd degree) and approximately 0.12 Hz for HF (83-point width, 3rd degree which was subtracted from the raw data to make a residual series (see Porges and Bohrer, 1990). This residual series was used to deflate the interpolated raw data proportional to the expected bias as in a), and the series was reverse-interpolated with the original time-series values to create a corrected series. In contrast to the above, this assumes nonstationarity and the final corrected series is hereafter referred to as NS-SPRV. It is unknown how this method of correction might interact with time-domain methods (SDNN and RMSSD), which were consequently not calculated. The metrics from Table 1 were re-calculated on the corrected series as above, to determine the improvements in dispersion. This is

20

Error (SPRV-ECG) (ms)

3.2. Experiment 1: pulse-to-sinus rhythm correction

10

0

-10

-20 700

800

900

1000

SPRV period (ms) Fig. 2. Beat length via the SPRV system in milliseconds (x-axis) vs. the corresponding error of ECG vs. SPRV comparison, with linear regression and 95% confidence interval displaying positive bias. ECG = electrocardiograph; SPRV = smartphone-mediated pulse rate variability.

J.A.J. Heathers / International Journal of Psychophysiology 89 (2013) 297–304 Table 2 Dispersion properties of ECG and SPRV values from Bland and Altman (1986) and correlation with error series. Bias Mean = Bland–Altman overall bias; Bias SD = standard deviation of Bland–Altman bias; LoA − ve = negative Bland–Altman Limit of Agreement; LoA + ve — Altman Limit of Agreement; recg-ERR = correlation between ECG and error series; rSPRV-ERR = correlation between SPRV and error series; ***p b 0.001, *p b 0.05. Participant

Bias mean

Bias SD

LoA − ve

LoA + ve

r (ECG-ERR)

r (SPRV-ERR)

1 2 3 4 5 6 7 8 9 10

0.47 0.56 0.48 0.32 0.24 0.32 0.32 0.32 0.42 0.36

4.60 4.33 2.68 6.81 5.40 5.46 6.41 4.43 6.44 3.71

−8.55 −7.92 −4.77 −13.02 −10.34 −10.38 −12.23 −8.36 −12.21 −6.92

9.48 9.05 5.74 13.66 10.81 11.02 12.87 9.00 13.05 7.63

0.232*** 0.048 0.55*** 0.375*** 0.488*** 0.259*** 0.216*** 0.146*** 0.567*** 0.245***

0.339*** 0.115* 0.578*** 0.512*** 0.568*** 0.347*** 0.285*** 0.24*** 0.606*** 0.35***

301

p b 0.001, and reduced both lnLF power (p = 0.032) and lnHF power (p b 0.004). The systematic positive bias for raw SPRV vs. ECG values during rest, attention and exercise (3 conditions) was again observed for all 5 HR and HRV variables(paired-samples t-tests; for smallest comparison: t9 = 3.11, p b 0.013). The errors for raw and corrected lnLF and lnHF values vs. ECG are shown for all conditions in Fig. 4. The proportion of the absolute values between these figures reveals the proportion of the initial percentage error which was normalised. Thus, linear correction to the S-SPRV series decreased SPRV-to-ECG error in lnLF at rest by 39%, but increased error during attention by 55%, and during exercise by 118%. S-SPRV correction decreased error in lnHF at all points (by 45%, 73%, 76%), and NS-SPRV decreased error for all values of lnLF (77%, 87%, 65%) and lnHF (95%, 81%, 66%).

3.4. Overall accuracy summarised in Fig. 3, which compares the percentage of error present on HR and common HRV metrics between methods. 3.3. Experiment 2: replication and task changes Experiment 2 replicated the close agreement between series found in Experiment 1 in all 30 recording periods. Correlation coefficients (Pearson's r) between RR and raw PP series were highest at rest (range: 0.993–0.997, median: 0.996) and during sustained attention, (range: 0.998–0.991, median: 0.995), and very slightly attenuated during exercise (range: 0.965–0.998, median: 0.989). The attentional task produced a typical moderate decrease in lnLF power (p = 0.038) but no change in HR (p = 0.13) or lnHF power (p = 0.46) as measured by ECG. Exercise reduced heart period as intended — rest 849.1 ms (50.3) vs. exercise 748.3 ms (43.3),

Fig. 5 displays the overall histogram of error present in all comparisons (Experiment 1: n = 10, 10 values; Experiment 2: n = 10, 30 values) and their corresponding QQ plots. Both error distributions are approximately Gaussian for both LF (n = 40, bias = 0.04%, SD = 0.61) and HF (n = 39, bias = 0.12%, SD = 2.1; n = 1 outlier was excluded). All values excluding the outlier fell overwhelmingly within +/− 2% (lnLF) or +/− 5% (lnHF) around the bias at near-zero. An outlier with a raw HF power of 9 ms2 (ECG) vs. 5 ms2 (SPRV) was excluded due to near-complete vagal withdrawal during task — the participant's mean heart rate during exercise of 497 ms (approximately 120BPM), above the likely intrinsic heart rate (Jose and Collison, 1970). Thus, the inaccuracy of this result is most likely the consequence of the simplicity in the basic exercise task used, which did not control for workload, baseline heart rate or calculate change in HR proportionally, rather than the inaccuracy of the SPRV system.

10

10 ln LF ln HF

% deviation from ECG

% deviation from ECG

5

0

5

0

LF

HF ln

ln

Fig. 3. SPRV correction method for individual HRV metrics (x-axis) vs. mean and 95% confidence intervals of error vs. equivalent ECG values (y-axis). ECG = electrocardiograph; SPRV = smartphone-mediated pulse rate variability; mean RR = average heart rate; SDNN = standard deviation of normal-to-normal intervals; RMSSD = root-mean squared of successive differences; ln LF = natural logarithm of low-frequency spectral power (0.04 to 0.15 Hz); ln HF = natural logarithm of low-frequency spectral power (0.04 to 0.15 Hz); RAW = uncorrected SPRV; S-SPRV = SPRV linearly corrected i.e. with the assumption of stationarity; NS-SPRV = SPRV corrected by residual polynomial series i.e. without the assumption of stationarity.

RAW

S-SPRV

EX

AT T

RE ST

EX

T AT

RE ST

EX

T AT

HF

LF ln

ln

RM SS D

Me

LF

HF ln

an RR SD NN

S-SPRV

-5 NS-SPRV

RE ST

RAW

ln

RM SS D

Me

an

RR SD NN

-5

NS-SPRV

Fig. 4. SPRV correction method for individual collection environments (x-axis) vs. mean and 95% confidence intervals of error vs. equivalent ECG values (y-axis), separate values for ln LF and ln HF. ECG = electrocardiograph; SPRV = smartphone-mediated pulse rate variability; REST = comparison at rest; ATT = comparison during attentional load; EX = comparison during exercise; RAW = uncorrected SPRV; S-SPRV = SPRV linearly corrected i.e. with the assumption of stationarity; NS-SPRV = SPRV corrected by residual polynomial series i.e. without the assumption of stationarity.

302

J.A.J. Heathers / International Journal of Psychophysiology 89 (2013) 297–304

-3

-2

4

4

2

2

-1

1

2

3 -3

-2

-1

1

-2

-2

-4

-4

rank-based z-score (LF)

2

3

rank-based z-score (HF)

80

LF HF Values (%)

60

40

20

0 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6

-6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6

Error from NS-SPRV to ECG (%) Fig. 5. Percentage error for the individual values given by the NS-SPRV correction for ln LF (x-axis, left), and ln HF (x-acis, right). Immediately above are corresponding QQ plots for ln LF (left) and ln HF (right). ECG = electrocardiograph; SPRV = smartphone-mediated pulse rate variability; NS-SPRV = SPRV corrected by residual polynomial series i.e. without the assumption of stationarity.

4. Discussion The SPRV system achieved a substantial level of accuracy at rest. The raw data and its analysis in both time and frequency domains correspond systematically to data from a simultaneous ECG. SPRV values display a systematic bias which precisely replicates that described by Giardino et al. (2002), which effectively corrected in Experiment 1 and separately verified in Experiment 2. At rest, a trivial linear correction for a normalised overall error managed to reduce error in some cases. A detrending method was more successful overall, reducing error as seen in Fig. 4. Task effects do not dramatically modify the accuracy of the SPRV system. As per rest, raw SPRV values were slightly but consistently larger for all metrics, and corrections to frequency analytic values were effective, with the residual detrended correction again more accurate. Accuracy was not drastically affected by task changes, but exercise tended to decrease the correlation between ECG and SPRV series, and affected the reliability of all calculated metrics. This may be due either to systematic cardiovascular changes during exercise, which has coordinated autonomic, cardiac and vascular effects that are difficult to separate, or due to problems with the measurement environment itself (i.e. movements of the electrodes or sensor due to exercise). Overall, SPRV accuracy compares favourably to that displayed by a variety of commercial sports monitors previously compared to analytical equipment (Leger and Thivierge, 1988; Terbizan et al., 2002; Weippert et al., 2010), and compares similarly well to analytical monitors co-compared with automatic filtering (Sandercock et al., 2004). The fact that SPRV values provide an approximation to the precise underlying cardiac period has to be balanced by the potential to greatly increase sample size through both remote and parallel data collection. There is no rule of thumb which might define an

acceptable criterion for SPRV accuracy out of context — instead, the measured accuracy must be balanced against the consideration of sample size, and the magnitude of the expected effect to be observed. A power analysis is instructive here: for example, if we take the reduction of ECG-derived LF-power during sustained attention compared to rest, and perform power analysis using the difference in the means and the pooled standard deviation above with α = 80%, this planned comparison would require n = 26 participants. If a hypothetical system which was significantly less accurate – say, with fully double the standard deviation of the analytical method – was used, then the required sample size rises to n = 95. With equipment capable of mass-sample collection, this would be a substantially more practical comparison to perform. That being said, if the accuracy of the SPRV system can be established, this represents an extreme case and there is also the potential that comparisons may be over-powered (i.e. that samples may be of such a size that practically or clinically insignificant effects may be statistically significant). However, this problem is well recognised and largely ameliorated by the reporting of effect size (Lenth, 2001). The pulse-ECG error in lnHF values, compared in Fig. 5, is clearly more substantial than lnLF in line with previous evidence for photoplethysmographic comparisons (Charlot et al., 2009; Giardino et al., 2002). This is of interest as HF and HF-related HRV metrics are more commonly researched dependent variables than LF, due to the relationship between the cardiovagal system and respiratory sinus arrhythmia (RSA). It stands to reason that HF power measured by pulse approximates respiratory sinus arrhythmia, but also potentially involves changes due to the modification of the local vasculature due to blood pressure modulation at respiratory frequency driven by changes in thoracic pressure. An interesting future comparison would be to characterise the accuracy of the present system simultaneously with an ear photoplethysmograph to differentiate

J.A.J. Heathers / International Journal of Psychophysiology 89 (2013) 297–304

between systemic and local vasoconstrictive responses; pulse measurement shows clear differences between ear and finger sites during vasoconstrictive challenge (Awad et al., 2001; Jablonka et al., 2008). This also highlights the need for caution when assessing comparative methods using normalised units. While this is a common calculation for frequency, and one of the few methods of standardising frequency output (which varies greatly in methodology between studies), it is unsuitable for characterising frequency differences between comparative methods. If at any point, the values of LF and HF assort differently (as they appear to here), then reminiscent of Eckberg's, 1997 criticism of sympathovagal balance, the proportional change in either LF or HF is obscured by their normalisation. It should also be noted that the correction method was proportional only to the relative amount of bias initially measured – this was calculated in Experiment 1 and not derived from individual data series in principle. In addition, the relationship between the mean cardiac cycle and the amount of raw or proportional error, for instance, was not added to the correction. It remains to be seen whether or not this result will hold in a more representative population sample. It is also likely that the accuracy of the system displayed here can be further improved. The correction parameters as derived in Experiment 1, while somewhat effective in a separate sample (Experiment 2), are specific to the population from which the sample is drawn — a more sophisticated method of correction might in future utilise the pulse height, cardiac cycle length, the magnitude of the derivative peak, or other vascular information available in the pulse signal to better characterise the pulse-ECG interval from first principles. Also, there is little agreement on which part of the photoplethysmographic waveform should be used to define the cardiac cycle, and there is no analytical exposition comparing the properties of either a pulse rate series or pulse transit time between the different definitions of cardiac cycles measured by pulse — fiducial points have been taken from the foot notch (Drinnan et al., 2001; Ma and Zhang, 2005), systolic peak (Charlot et al., 2009), a percentage threshold of the signal size (Smith et al., 1999), and first derivative peak (Heathers et al., 2012). Considering that crest time (the time between the foot and peak of the ‘up’ section of the pulse beat) is one of the classic components of the pulsatile waveform related to vascular disease (Dillon and Hertzman, 1941), the values of these points may change relative to each other depending on the local or global vascular environment. Assuming accuracy can be maintained or improved, the intended SPRV implementation is simple — a sensor is provided for analysis and the software required is downloaded to a smartphone (or equivalent device), which is provided by the participants. Alternatively, in face-to-face research, a device may be provisioned then re-used. After collection, data is emailed directly to experimenters from the device. This not only has the potential to increase psychophysiological recording capacity, but lowers the need for the experimenter's direct involvement. If sensors can be shipped, there are possibilities for remote or rural data collection, naturalistic monitoring (i.e. during a stressful period), time-of-day monitoring during periods normally hard to access (i.e. immediately after waking) or access to broadly dispersed populations (i.e. sufferers of an uncommon malady). The cost of the sensor may also not necessitate its retrieval. This work has some limitations, the most central of which is that the above calculations are performed to a small sample of ideal data — collected under supervision, repeated or extended if necessary, and hand-corrected. The present work does not address the various issues in and around the correction of common sources of error such as ectopy or movement. Optical pulse systems are sensitive to movement artifacts and in a normal collection environment necessitate an analytical approach to identifying and correcting for errors. Considering the adaptive filtering approach used, the error most commonly seen in the SPRV system is a ‘skipped’ beat rather than an overdetection. The interaction between beat replacement technique

303

and spectral analysis is well known (Clifford and Tarassenko, 2005), and also warrants investigation. Two further caveats are necessary. Firstly, the pulsatile waveform differs in both its hemodynamic properties (Awad et al., 2007) and its typical displayed spectra (Allen, 2007) depending on the analysis site. The results, and especially the correction methods in the present work, should be considered confined to the pulse signal at the finger rather than the ear or toe. Secondly, due to the relationship between vascular health and age (Millasseau et al., 2002) it seems prudent that the stated accuracy be limited to healthy young adults. 5. Conclusion The SPRV system records an approximation of an ECG-derived RRseries sufficient to accurately perform time- and frequency-domain analysis of its variability, and accurately reflects change in autonomic output provided by typical psychophysiological stimuli. Furthermore, systematic circulatory and vascular factors allow the SPRV signal, without reference to a baseline signal, to be significantly corrected for estimated error. This system has the potential to expand broad-based, remote or mass experimental HR data collection, and has significant advantages in accessibility, portability and cost over traditional laboratory systems for the provision of cardiac psychophysiological data. Supplementary data to this article can be found online at http:// dx.doi.org/10.1016/j.ijpsycho.2013.05.017. Acknowledgements The author wishes to acknowledge the superlative engineering and technical assistance of Simon Wegerif, and the valuable insight and suggestions of Dr. Marc de Rosnay into the final manuscript. References Aasman, J., Mulder, G., Mulder, L.J.M., 1987. Operator effort and the measurement of heart-rate variability. Human Factors 29, 161–170. Allen, J., 2007. Photoplethysmography and its application in clinical physiological measurement. Physiological Measurement 28, R1–R39. Awad, A.A., Ghobashy, M.A.M., Ouda, W., Stout, R.G., Silverman, D.G., Shelley, K.H., 2001. Different responses of ear and finger pulse oximeter wave form to cold pressor test. Anesthesia & Analgesia 92, 1483–1486. Awad, A.A., Haddadin, A.S., Tantawy, H., Badr, T.M., Stout, R.G., Silverman, D.G., Shelley, K.H., 2007. The relationship between the photoplethysmographic waveform and systemic vascular resistance. Journal of Clinical Monitoring and Computing 21, 365–372. Bernardi, L., Wdowczyk-Szulc, J., Valenti, C., Castoldi, S., Passino, C., Spadacini, G., Sleight, P., 2000. Effects of controlled breathing, mental activity and mental stress with or without verbalization on heart rate variability. Journal of the American College of Cardiology 35, 1462–1469. Berntson, G.G., Stowell, J.R., 1998. ECG artifacts and heart period variability: don't miss a beat! Psychophysiology 35, 127–132. Blackman, R.B., Tukey, J.W., 1959. The Measurement of Power Spectra from the Point of View of Communication Engineering. Dover Publications, New York. Bland, J.M., Altman, D.G., 1986. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 8476, 307–310. Bland, J.M., Altman, D.G., 1995. Comparing methods of measurement — why plotting difference against standard method is misleading. Lancet 8982, 1085–1087. Bolanos, M., Nazeran, H., Haltiwanger, E., 2006. Comparison of heart rate variability signal features derived from electrocardiography and photoplethysmography in healthy individuals. IEEE Eng Med Bio, pp. 4289–4294. Charlot, K., Cornolo, J., Brugniaux, J.V., Richalet, J.P., Pichon, A., 2009. Interchangeability between heart rate and photoplethysmography variabilities during sympathetic stimulations. Physiological Measurement 30, 1357–1369. Clifford, G.D., Tarassenko, L., 2005. Quantifying errors in spectral estimates of HRV due to beat replacement and resampling. IEEE Transactions on Biomedical Engineering 5252, 630–638. Deditius-Island, H.K., Caruso, J.C., 2002. An examination of the reliability of scores from Zuckerman's Sensation Seeking Scales, Form V. Educational and Psychological Measurement 62, 728–734. Denver, J.W., Reed, S.F., Porges, S.W., 2007. Methodological issues in the quantification of respiratory sinus arrhythmia. Biological Psychology 74, 286–294. Dillon, J.B., Hertzman, A.B., 1941. The form of the volume pulse in the finger pad in health, atherosclerosis, and hypertension. American Heart Journal 21, 172–190. Drinnan, M.J., Allen, J., Murray, A., 2001. Relation between heart rate and pulse transit time during paced respiration. Physiological Measurement 22, 425–432.

304

J.A.J. Heathers / International Journal of Psychophysiology 89 (2013) 297–304

Eckberg, D.L., 1997. Sympathovagal balance: a critical appraisal. Circulation 96, 3224–3232. Foo, J.Y., Wilson, S.J., Williams, G.R., Harris, M.A., Cooper, D.M., 2005. Pulse transit time changes observed with different limb positions. Physiological Measurement 26, 1093–1102. Free, C., Phillips, G., Galli, L., Watson, L., Felix, L., Edwards, P., Patel, V., Haines, A., 2013a. The effectiveness of mobile-health technology-based health behaviour change or disease management interventions for health care consumers: a systematic review. PLoS Medicine (published 15/01/2013). Free, C., Phillips, G., Watson, L., Galli, L., Felix, L., Edwards, P., Patel, V., Haines, A., 2013b. The effectiveness of mobile-health technologies to improve health care service delivery processes: a systematic review and meta-analysis. PLoS Medicine (published 15/01/13). Giardino, N.D., Lehrer, P.M., Edelberg, R., 2002. Comparison of finger plethysmograph to ECG in the measurement of heart rate variability. Psychophysiology 39, 246–253. Gil, E., Orini, M., Bailon, R., Vergara, M., Mainardi, L., Laguna, P., 2010. Photoplethysmography pulse rate variability as a surrogate measurement of heart rate variability during non-stationary conditions. Physiological Measurement 31, 1271–1290. Goldberger, A.L., Peng, C.K., Lipsitz, L.A., 2002. What is physiologic complexity and how does it change with aging and disease? Neurobiology of Aging 23, 23–26. Grippo, A.J., Johnson, A.K., 2002. Biological mechanisms in the relationship between depression and heart disease. Neuroscience and Biobehavioral Reviews 26, 941–962. Grossman, P., Taylor, E.W., 2007. Toward understanding respiratory sinus arrhythmia: relations to cardiac vagal tone, evolution and biobehavioral functions. Biological Psychology 74, 263–285. Heathers, J.A., Fink, E., Kuhnert, R.L., de Rosnay, M., 2012. Blood volume pulse (BVP) derived vagal tone (VT) between 5 and 7 years of age: a methodological investigation of measurement and longitudinal stability. Developmental Psychobiology (Nov. 5th [Epub ahead of print]). Hertzman, A.B., 1938. The blood supply of various skin areas as estimated by the photoelectric plethysmograh. American Journal of Physiology 124, 328–340. Jablonka, D.H., Awad, A.A., Stout, R.G., Silverman, D.G., Shelley, K.H., 2008. Comparing the effect of arginine vasopressin on ear and finger photoplethysmography. Journal of Clinical Anesthesia 20, 90–93. Jose, A.D., Collison, D., 1970. The normal range and determinants of the intrinsic heart rate in man. Cardiovascular Research 4, 160–167. Leger, L., Thivierge, M., 1988. Heart rate monitors: validity, stability, and functionality. The Physician and Sportsmedicine 16, 143–148. Lenth, R.V., 2001. Some practical guidelines for effective sample-size determination. The American Statistician 55, 187–193. Lu, C.L., Zou, X., Orr, W.C., Chen, J.D., 1999. Postprandial changes of sympathovagal balance measured by heart rate variability. Digestive Diseases and Sciences 44, 857–861. Lu, G., Yang, F., Taylor, J.A., Stein, J.F., 2009. A comparison of photoplethysmography and ECG recording to analyse heart rate variability in healthy subjects. Journal of Medical Engineering & Technology 33, 634–641. Ma, T., Zhang, Y.T., 2005. A correlation study on the variabilities in pulse transit time, blood pressure, and heart rate recorded simultaneously from healthy subjects. IEEE Eng Med Bio, pp. 996–999. Meijman, T.F., 1997. Mental fatigue and the efficiency of information processing in relation to work times. International Journal of Industrial Ergonomics 20, 31–38. Middleton, H.C., Sharma, A., Agouzou, D., Sahakian, B.J., Robbins, T.W., 1999. Contrasts between the cardiovascular concomitants of tests of planning and attention. Psychophysiology 36, 610–618. Millasseau, S.C., Kelly, R.P., Ritter, J.M., Chowienczyk, P.J., 2002. Determination of agerelated increases in large artery stiffness by digital pulse contour analysis. Clinical Science (London, England) 103, 371–377. Mulder, L.J.M., Mulder, G., 1987. Cardiovascular reactivity and mental work-load. In: Kitney, R.I., Rompelman, O. (Eds.), The Beat-to-beat Investigation of Cardiovascular Function: Measurement, Analysis, and Application. Clarendon, Oxford, UK, pp. 216–253. Mulder, B., Veldman, H., van der Veen, F., van Roon, A., Rüddel, H., Schächinger, H., Mulder, B., 1993. On the effects of mental task performance on heart rate, blood pressure and its variability measures. In: Di Rienzo, M., Mancia, G., Parati, G., Pedotti, A., Zanchetti, A. (Eds.), Blood Pressure and Heart Rate Variability: Computer Analysis, Modelling and Clinical Applications. IOS Press, Fairfax, VA, pp. 153–167.

Nickel, P., Nachreiner, F., 2003. Sensitivity and diagnosticity of the 0.1-Hz component of heart rate variability as an indicator of mental workload. Human Factors 45, 575–590. Obrist, P.A., Light, K.C., McCubbin, J.A., Hutcheson, J.S., Hoffer, J.L., 1979. Pulse transit time: relationship to blood pressure and myocardial performance. Psychophysiology 16, 292–301. Ode, S., Hilmert, C.J., Zielke, D.J., Robinson, M.D., 2010. Neuroticism's importance in understanding the correlates of heart rate variability: a daily life outcome study of stress, distress, and dysfunction. Emotion 10, 536–543. Payne, R.A., Symeonides, C.N., Webb, D.J., Maxwell, S.R., 2006. Pulse transit time measured from the ECG: an unreliable marker of beat-to-beat blood pressure. Journal of Applied Physiology 100, 136–141. Pew Research Center's Internet & American Life Project, 2012. http://pewinternet.org/ Reports/2012/Smartphone-Update-2012.aspx (Accessed 20/01/13). Pinna, G.D., Maestri, R., Torunski, A., Danilowicz-Szymanowicz, L., Szwoch, M., La Rovere, M.T., Raczak, G., 2007. Heart rate variability measures: a fresh look at reliability. Clinical Science (London, England) 113, 131–140. Pitzalis, M.V., Mastropasqua, F., Massari, F., Forleo, C., Di Maggio, M., Passantino, A., Colombo, R., Di Biase, M., Rizzon, P., 1996. Short- and long-term reproducibility of time and frequency domain heart rate variability measurements in normal subjects. Cardiovascular Research 32, 226–233. Poon, C.C.Y., Zhang, Y.T., 2006. Cuff-less and noninvasive measurements of arterial blood pressure by pulse transit time. 27th Conf Proc IEEE Eng Med Biol Soc, pp. 5877–5880. Porges, S.W., Bohrer, R.E., 1990. The analysis of periodic processes in psychophysiolical research. In: Cacioppo, J.T., Tassinary, L.G. (Eds.), Principles of Psychophysiology: Physical, Social and Inferential Elements. Cambridge University Press, New York, NY, pp. 708–753. Routledge, H.C., Chowdhary, S., Coote, J.H., Townend, J.N., 2002. Cardiac vagal response to water ingestion in normal human subjects. Clinical Science (London, England) 103, 157–162. Sandercock, G., 2007. Normative values, reliability and sample size estimates in heart rate variability. Clinical Science (London, England) 113, 129–130. Sandercock, G.R., Brodie, D.A., 2006. The use of heart rate variability measures to assess autonomic control during exercise. Scandinavian Journal of Medicine & Science in Sports 16, 302–313. Sandercock, G.R., Bromley, P.D., Brodie, D.A., 2004. Reliability of three commercially available heart rate variability instruments using short-term (5-min) recordings. Clinical Physiology and Functional Imaging 24, 359–367. Sandercock, G.R., Bromley, P.D., Brodie, D.A., 2005. The reliability of short-term measurements of heart rate variability. International Journal of Cardiology 103, 238–247. Simmons, J., Nelson, L., Simonsohn, U., 2011. False-positive psychology: undisclosed flexibility in data collection and analysis allow presenting anything as significant. Psychological Science 22, 1359–1366. Smith, R., Argod, J., Pepin, J., Levy, P., 1999. Pulse transit time: an appraisal of potential clinical applications. Thorax 54, 452–457. Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology, 1996. Heart rate variability. Standards of measurement, physiological interpretation, and clinical use. European Heart Journal 17, 354–381. Terbizan, D.J., Dolezal, B.A., Albano, C., 2002. Validity of seven commercially available heart rate monitors. Measurement in Physical Education and Exercise Science 6, 243–247. Thayer, J.F., Åhs, F., Fredrikson, M., Sollers, J.J., Wager, T.D., 2012. A meta-analysis of heart rate variability and neuroimaging studies: implications for heart rate variability as a marker of stress and health. Neuroscience and Biobehavioral Reviews 36, 747–756. Weippert, M., Kumar, M., Kreuzfeld, S., Arndt, D., Rieger, A., Stoll, R., 2010. Comparison of three mobile devices for measuring R–R intervals and heart rate variability: Polar S810i, Suunto t6 and an ambulatory ECG system. European Journal of Applied Physiology 109, 779–786. Williams, E.J., 1959. The comparison of regression variables. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 21, 396–399. Wong, J.-S., Lu, W.-A., Wu, K.-T., Liu, M., Chen, G.-Y., Kuo, C.-D., 2012. A comparative study of pulse rate variability and heart rate variability in healthy subjects. Journal of Clinical Monitoring and Computing 26, 107–114. Zong, W., Heldt, T., Moody, G.B., Mark, R.G., 2003. An open-source algorithm to detect onset of arterial blood pressure pulses. Computers in Cardiology 30, 259–262.