International Journal of Cardiology 166 (2013) 15–29
`
`Contents lists available at SciVerse ScienceDirect
`
`International Journal of Cardiology
`
`j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / i j c a r d
`
`Review
`How accurate is pulse rate variability as an estimate of heart rate variability?
`A review on studies comparing photoplethysmographic technology
`with an electrocardiogram
`Axel Schäfer a,b, Jan Vagedes a,c,⁎
`a Arcim institute, Im Haberschlai 7, 70794 Filderstadt, Germany
`b University of Tübingen, Institute of Medical Psychology and Behavioral Neurobiology, Gartenstr. 29, 72074 Tübingen, Germany
`c University of Tübingen, Children's Hospital, Department of Neonatology, Calwerstrasse 7, 72076 Tübingen, Germany
`
`a r t i c l e
`
`i n f o
`
`a b s t r a c t
`
`Article history:
`Received 7 September 2011
`Received in revised form 26 January 2012
`Accepted 10 March 2012
`Available online 17 July 2012
`
`Keywords:
`Heart rate variability
`Photoplethysmography
`Pulse wave analysis
`Continuous blood pressure monitoring
`Electrocardiography
`
`Background: The usefulness of heart rate variability (HRV) as a clinical research and diagnostic tool has been
`verified in numerous studies. The gold standard technique comprises analyzing time series of RR intervals
`from an electrocardiographic signal. However, some authors have used pulse cycle intervals instead of RR
`intervals, as they can be determined from a pulse wave (e.g. a photoplethysmographic) signal. This option
`is often called pulse rate variability (PRV), and utilizing it could expand the serviceability of pulse oximeters
`or simplify ambulatory monitoring of HRV.
`Methods: We review studies investigating the accuracy of PRV as an estimate of HRV, regardless of the under-
`lying technology (photoplethysmography, continuous blood pressure monitoring or Finapresi, impedance
`plethysmography).
`Results/conclusions: Results speak in favor of sufficient accuracy when subjects are at rest, although many
`studies suggest that short-term variability is somewhat overestimated by PRV, which reflects coupling effects
`between respiration and the cardiovascular system. Physical activity and some mental stressors seem to
`impair the agreement of PRV and HRV, often to an inacceptable extent. Findings regarding the position of
`the sensor or the detection algorithm are not conclusive.
`Generally, quantitative conclusions are impeded by the fact that results of different studies are mostly incom-
`mensurable due to diverse experimental settings and/or methods of analysis.
`© 2012 Published by Elsevier Ireland Ltd.
`
`1. Introduction
`
`The term ‘Heart rate variability’ (HRV) refers to the fact that the
`duration of cardiac cycles is not constant, but varies from one heart-
`beat to the next. The extent of variability is determined by digital
`processing of an electrocardiographic (ECG) signal. Because of their
`distinct profile, the R peaks of an ECG signal are suitable for automated
`detection by computer algorithms; hence the standard method to
`assess cardiac cycles is to place their limits at the R peaks. As a result,
`one obtains a time series of such consecutive RR intervals. Ectopic
`beats and arrhythmic events are usually not processed when deter-
`mining HRV; only regular beats should be considered, which is why
`one often encounters the alternative term NN intervals (‘normal to
`normal’). Analysis of HRV comprises the computing of meaningful
`parameters from RR interval time series called HRV variables. For a
`comprehensive presentation of the methodology see [1].
`
`⁎ Corresponding author at: University of Tübingen, Children's Hospital, Department
`of Neonatology, Calwerstrasse 7, 72076 Tübingen, Germany. Tel.: +49 7071 2984742.
`E-mail address: Jan.Vagedes@med.uni-tuebingen.de (J. Vagedes).
`
`0167-5273/$ – see front matter © 2012 Published by Elsevier Ireland Ltd.
`doi:10.1016/j.ijcard.2012.03.119
`
`HRV has become a very useful tool in clinical diagnostics within
`the last decades. Reduced HRV is correlated with the risk of cardiac
`events like myocardial infarction and congestive heart failure [2,3],
`and with sudden cardiac death [4,5]. From the early 1980s on, the
`frequency domain HRV variables, gained from a power spectrum of
`the RR interval series, have been found to reflect autonomic cardio-
`vascular control [6,7]. Physical fitness and social integration are both as-
`sociated with reduced cardiac risk and enhanced HRV [8,9]. Treatment
`of psychiatric patients with tricyclic antidepressants seriously dimin-
`ishes HRV, although there is ambiguity as to whether psychopharmaco-
`logical treatment increases mortality [10]. Also, depression itself seems
`to influence autonomic control and HRV [11]. All in all, HRV seems to
`be a reliable and multifunctional parameter indicating cardiovascular
`and autonomic health as well as general psychic and somatic fitness.
`Evaluation of HRV variables can be gained from sessions of various
`duration, up to Holter recordings lasting 24 h or more. However,
`short-term recordings of only a few minutes have been found to be
`comparably useful [12], and even ultra-short sequences of only 10 s
`seem to have reasonable diagnostic value [13].
`Photoplethysmography (PPG) is a technique developed in the 1930s
`for monitoring blood volume changes in the micro vascular bed of tissue
`
`APPLE 1018
`
`1
`
`

`

`16
`
`A. Schäfer, J. Vagedes / International Journal of Cardiology 166 (2013) 15–29
`
`[14]. In more recent decades, progress in semiconductor technology and
`optoelectronics, as well as advancements in digital signal processing,
`have facilitated a renaissance of PPG, which today is probably the
`most widespread method used in clinical monitoring. Its basic principle
`requires a light source to illuminate subcutaneous tissue (typically
`an LED, i.e. a light emitting diode) and a photo detector with spectral
`characteristics matching those of the light source (e.g. a photodiode
`or phototransistor). Current PPG sensors use low-cost optoelectronic
`components operating in the domain of red and/or near infrared
`wavelengths.
`There are two basic configurations used in PPG: transmission
`mode, where the perfuse tissue (like a fingertip or an earlobe) is
`placed between the source and the detector, and reflection mode,
`where the two electronic components are placed side-by-side near
`the skin, e.g. at the forehead. In both cases the detector registers
`small variations in the transmitted or reflected light, respectively,
`caused by changes in microcirculation. Major factors affecting the
`detected light intensity are blood volume, blood vessel wall move-
`ment and the orientation of erythrocytes [15].
`Changes in a PPG signal arise from variations both in the path-
`length between source and detector and in the optical density of
`the blood. The signal can be decomposed into two parts [16]. The
`small pulsatile component, or AC component, arises from arterial
`blood pulsation; hence its oscillation parallels momentary cardiac
`activity. It is superimposed on the much larger DC component,
`where DC refers to direct current, suggesting a static behavior. How-
`ever, the DC component is not entirely static, but includes variations
`slower than the heart rate due to venous volume fluctuations, vaso-
`motor activity and thermoregulation. Changes in the intrathoracic
`pressure due to respiration cause fluctuations in the venous return
`to the heart, which in turn modulates cardiac output and blood pres-
`sure. Ventilation thus induces fluctuation of both AC and DC compo-
`nents, which enables one to monitor respiratory activity by filtering
`and processing a PPG signal appropriately [17,18].
`PPG technology is a very versatile tool, and its usefulness in a wide
`range of clinical applications has been demonstrated. The most
`common one is pulse oximetry, which utilizes the difference of red
`and near infrared light absorption by oxyhemoglobin and reduced
`hemoglobin to estimate arterial blood oxygen saturation. Further
`applications include the estimation of cardiac output, the diagnosis
`of atherosclerosis, peripheral arterial occlusion and other peripheral
`vascular diseases, as well as the assessment of arterial compliance
`and aging, of endothelial and venous function, micro vascular blood
`flow and other functions [15]. Additionally, PPG offers a number of
`ways to assess cardiovascular autonomic function and skin vasomotor
`function. Due to this property, it is also a valuable instrument for
`monitoring nociception during general anesthesia [16]. For a compre-
`hensive review on PPG and its clinical applications see [15].
`Arterial blood pressure can also be estimated from a PPG, because
`the higher the pressure, the quicker the propagation of the pulse
`wave from the heart to the periphery. Therefore, the determination
`of pulse wave velocity or its reciprocal, pulse transit time (cf.
`Section 3.1 below) for each pulse cycle can serve as an approximation
`of continuous blood pressure (CBP) monitoring. However, estimators
`of systolic and diastolic blood pressure are only in moderate agree-
`ment with their directly measured counterparts, although this can
`be somewhat improved by including the amplitude of each PPG
`pulse wave in the analysis [19,20]. In order to provide a more reliable
`method for CBP monitoring, the Finapres™ (for FINger Arterial PRES-
`sure) technology was introduced in the early 1980s. It is based on the
`dynamic vascular unloading of the arterial walls of the finger using an
`inflatable finger cuff [21]. The built-in PPG sensor is used as a control
`sensor to regulate cuff pressure for optimum CBP detection, but does
`not measure blood pressure directly.
`Given that PPG is such a simple and ubiquitous technology in clin-
`ical monitoring, it is preferable to maximize its potential. As already
`
`indicated above, the signal can also be used to monitor the heart
`and respiratory rate of a patient, including instantaneous rates or
`cycle lengths. In principle, this would allow HRV variables to be deter-
`mined as well from a PPG signal. Compared to the current standard of
`an ECG-based HRV analysis, this would involve certain benefits. In
`clinical situations where a pulse oximeter (PO) device is already at
`hand by default, being able to include HRV analysis in the monitoring
`process without requiring an ECG means a significant advantage. In
`addition, during magnetic resonance imaging (MRI), for example,
`ECG electrodes or other metal-containing sensors are not permitted,
`as they interfere with strong electromagnetic fields.
`Another fact speaking in favor of PPG technology is that it is
`noninvasive, cost-effective and straightforward to use. Detecting the
`signal usually requires no more than attaching a single sensor to a
`finger or an earlobe, compared to at least three leads and Ag/AgCl
`electrodes required for an ECG. Furthermore, ECG electrodes often
`have to be applied to the chest, requiring the patients to undress—
`which can delay recordings and pose a problem for embarrassed
`patients. On the other hand, a major disadvantage of PPG technology
`is that the signal is susceptible to motion artifacts, which can impair
`the accuracy of the detected cardiac activity [15].
`
`2. Scope of this review
`
`The purpose of this review is to summarize the hitherto existing
`literature about the accuracy of estimating HRV and/or (instanta-
`neous) heart rates from a continuously recorded pulse wave signal.
`All of the considered articles contrast the results of the latter with
`the gold standard of an ECG-based method.
`We performed a search in PubMed and Embase for publications
`that matched at least one keyword in both of the following word
`groups:
`
`1) oximetr…/oxymetr… or plethysmogr… or “pulse wave” or “pulse
`applan…”
`2) electrocard… or ECG or “rate variability”
`
`Here the ellipsis (…) indicates arbitrary completion of words, and
`quotation marks require the enclosed words to appear exactly in
`the specified order. The retrieved abstracts were then inspected for
`their thematic agreement with the topic of this review. Such selected
`articles and correspondingly eligible references cited therein were all
`included.
`Most of the retrieved studies use a PPG signal to detect a pulse
`wave. However, some employ a Finapres™ system monitoring CBP
`or an impedance plethysmography system instead. The analysis
`techniques used to discern individual pulse cycles in these cases are
`similar to those utilized for PPG signals. Additionally, many of the
`findings therein complement those in PPG studies. For these reasons,
`all those references assessing the quality of results from non-PPG
`pulse waves are included in this review.
`The main focus is on the question whether the ECG-based method
`of evaluating HRV can be replaced by a technique using a pulse wave.
`From now on the latter approach will be referred to as pulse rate
`variability (PRV), as it is based on the varying length of pulse cycles,
`not cardiac cycles. Heartbeats can be more accurately determined
`from an ECG; therefore we will henceforth use the term HRV mainly
`for ECG-derived heart rate variability.
`It is clear that the methods used to determine the limits between
`adjacent pulse cycles affect the individual lengths of the latter and,
`hence, PRV results. The statistical techniques used to compare, say
`PRV and HRV, will likewise have an impact on the outcomes of an
`investigation. Section 3 outlines some important approaches used in
`analyses; additionally, Tables 1 and 2 summarize the most important
`abbreviations regarding HRV variables and HRV/PRV analysis.
`A number of researchers were merely interested in the accuracy of
`estimating the average heart rate with a PPG device for monitoring
`
`2
`
`

`

`A. Schäfer, J. Vagedes / International Journal of Cardiology 166 (2013) 15–29
`
`17
`
`Table 1
`Abbreviations of common HRV variables.
`
`Time domain variables
`
`Mean NN, mean HR
`
`SDNN
`
`RMSSD
`
`SDSD
`
`NN50/pNN50 (pNN10)
`
`Mean of all NN intervals. Mean HR is its reciprocal value,
`converted into beats per minute (bpm)
`Standard deviation of all NN intervals;
`measure of overall variability.
`Root of mean of squared subsequent differences;
`measure of short-term variability.
`Standard deviation of subsequent differences
`(almost identical to RMSSD).
`Number/percent of subsequent differences with an
`absolute value > 50 ms (or 10 ms).
`
`Frequency domain variables (spectral domains according to Task force definitions [1])
`
`ULF
`VLF
`LF
`HF
`TP
`
`Nonlinear variables
`
`SD1
`SD2
`ApEn
`SampEn
`
`Ultra low frequencies (b0.003 Hz)
`Very low frequencies (0.003–0.04 Hz)
`Low frequencies (0.04–0.15 Hz)
`High frequencies (0.15–0.40 Hz)
`Total power (sum contribution of all spectral domains)
`
`Standard deviation of short diagonal axis in Poincaré plot
`Standard deviation of long diagonal axis in Poincaré plot
`Approximate entropy
`Sample entropy
`
`purposes; hence they did not concern themselves with individual
`heart or pulse cycles or HRV variables. However, some of them use
`analytical approaches that are interesting in context, while other in-
`vestigations can be perceived as historical predecessors of the studies
`in the focus of this review. We thus summarize these results briefly in
`Section 4 and Table 3.
`The central topic, compatibility of PRV with HRV, is treated in
`Section 5 and its subsections. We should note that the various find-
`ings are almost never comparable directly, since a variety of devices,
`analytical methods, experimental conditions etc. have been used.
`Therefore, we can present and contrast all results only descriptively,
`itemized in Tables 4 and 5.
`A discussion and conclusions can be found in Sections 6 and 7.
`
`3. Analytical methods
`
`3.1. Pulse wave analysis
`
`A typical pulse wave cycle can be subdivided into two parts (cf.
`Fig. 1). The anacrotic phase is the rising part of the pulse due to systole.
`Shortly after the QRS complex appears in the ECG, the ventricular systole
`generates a pulse wave travelling distally. In the arteries and arterioles
`this leads to a rapid increase in blood pressure and blood volume, i.e. a
`steep rise in the pulse wave. The subsequent decline corresponding to
`cardiac diastole is termed the catacrotic phase and is more prolonged
`than the anacrotic phase. Often it contains a secondary peak separated
`
`Table 2
`Abbreviations of HRV and PRV phenomena used in the text.
`
`Time intervals
`RRI
`PPI
`
`Frequencies/rates
`HR
`PR
`IHR
`IPR
`
`Result of analysis
`HRV
`PRV
`
`RR interval length(s) from an ECG
`Pulse interval length(s) from a pulse wave signal
`
`Heart rate, determined by an ECG signal
`Pulse rate, determined by a pulse wave signal
`Instantaneous heart rate (reciprocal of RRI)
`Instantaneous pulse rate (reciprocal of PPI)
`
`Heart rate variability, gained by analyzing RRI values
`Pulse rate variability, gained by analyzing PPI values
`
`by the so-called dicrotic notch, an effect diminishing with ageing and
`increasing arterial stiffness [15]. This phenomenon is often attributed
`to the closure of the aortic valve [22], although recent findings suggest
`a causation by reflection of the pressure peak from small arteries in
`the trunk and lower limbs [23].
`Corresponding to each RR interval (RRI) of the ECG, which is
`considered as the “true” instantaneous heart cycle length, there is
`an interval comprising a full pulse cycle length. We will henceforth
`denote it as PPI (“pulse to pulse interval”). Its exact location depends
`on the definition of its boundaries and the computer algorithm used
`to detect them [24]. The black disks in Fig. 1 highlight three possible
`alternatives. One can determine the beginning of the anacrotic or,
`alternatively, the catacrotic phase, i.e. the pulse foot or peak (marked
`with an “f” or “p” in the figure), as the respective boundary. This leads
`to different definitions of pulse intervals, as is indicated in the figure
`by PPI (f) and PPI (p). A third option is to use the maximum 1st deriv-
`ative representing the steepest part of the upstroke as a boundary, in-
`dicated by the disk marked “d”. In the existing literature researchers
`have utilized all of these methods as well as others (cf. Table 5); com-
`parative studies on some approaches can be found in [25,26].
`Depending on the pulse wave velocity and the vascular path from
`the heart to the location of the detector, there is a delay between each
`R peak and the onset of its corresponding pulse wave. The delay is
`usually termed pulse transit time (PTT) and is negatively correlated
`with blood pressure, arterial stiffness, and age [15]. As defined
`above, it ranges from an R peak to the next pulse foot or the beginning
`of systole; however, many studies employ a different definition of a
`PTT extending to the subsequent peak. The two cases are denoted
`as PTT (f) and PTT (p) in Fig. 1.
`Deviations of the PPI from the RRI series can arise from two possi-
`ble causes: 1) an inaccurate detection of pulse cycle boundaries due
`to hardware limitations, artifacts and/or noise, or 2) a physiological
`variability in PTT.
`
`3.2. Heart and pulse rate variability analysis
`
`In order to quantify the agreement of PPI and RRI series, researchers
`have mostly used one or more of the following three strategies:
`
`a) Often one is only interested in the accuracy of the determined
`mean pulse rate (PR) compared to the mean heart rate derived
`from an ECG. This is usually sufficient when a PPG device is used
`for monitoring or telecare.
`b) The focus of this review is on the reliability of PRV as a substitute
`for HRV. Most of the studies in this domain compare HRV variables
`for both the RRI and PPI series, which we will henceforth refer to
`as HRV and PRV variables. The most common HRV variables eval-
`uated are time and frequency domain variables, which have been
`extensively standardized, but nonlinear variables are also used [1].
`Abbreviations of the most commonly used HRV variables are listed
`in Table 1.
`c) A number of authors also compare the RRI and PPI series directly,
`i.e. they compare the lengths of individual heart beats to the cor-
`responding pulse cycles—or they do the same thing with the in-
`stantaneous heart (IHR) and pulse (IPR) rates, which are simply
`the reciprocal values of the cycle lengths.
`
`Table 2 gives an overview of the respective quantities and their
`abbreviations used throughout the subsequent sections.
`
`3.3. Statistical analysis
`
`In all the studies considered in this review, quantities determined
`from an ECG and a simultaneously recorded pulse signal are com-
`pared. Data pairs are given where each pair comprises corresponding
`values measured by each method in question, respectively. In case
`a) of the preceding subsection, these values are average HR and PR
`
`3
`
`

`

`18
`
`A. Schäfer, J. Vagedes / International Journal of Cardiology 166 (2013) 15–29
`
`Table 3
`Overview of studies on monitoring heart rate with PPG technology.
`
`Study
`
`N=number of subjects×
`T=duration of records
`
`Device(s) used
`
`Position of
`sensor
`
`Experimental
`conditions
`
`Methods of pulse detection
`and analysis
`
`Results
`
`Adults
`Altemeyer et al.
`(1986) [31]
`
`Atlasz et al.
`(2006) [37]
`
`Lindberg et al.
`(1992) [34]
`Nakajima et al.
`(1996) [35]
`
`(Monitored patients
`during operation,
`ages 25–84) N = 19
`
`Biox III, and
`Nellcor N-100
`pulse oximeters
`
`Index or middle
`finger
`
`Both normal
`respiration and
`inhalation of hypoxic
`gas mixture.
`5 min supine rest,
`5 min standing
`
`Fingertip
`
`(Healthy, ages 22.2±1.9)
`
`N = 35 × T = 10 min
`(Males, ages 20–30)
`N = 11 × T = 10 min
`(Males, ages 22–34)
`N = 11 × T = 20 min
`
`Self-made
`reflection PPG
`sensor
`Self-made
`PPG sensor
`Self-made
`PPG sensor
`
`Left forearm
`
`Supine rest
`
`Earlobe
`
`Ergometer: sitting +
`Exercise (30–130 W)
`
`“Good agreement” with ECG-
`determined HR, not quantita-
`tively specified.
`
`No significant differences
`between HR from PPG and ECG.
`
`Exact agreement with HR from
`ECG, when signal clean
`PPG and ECG determined heart
`rates “agreed well”, maximum
`|ΔHR| = 10 bpm
`
`PCC of 0.95, Bland-Altman
`analysis says 95% CI: ΔHR =
`−0.1 ± 6.5 bpm
`
`Spectral method and infrared
`light both more reliable. Results
`in this case: PCC = 0.92, ΔHR =
`−0.6 ± 1.3 bpm (mean ± SD)
`
`Auto-detection method of
`monitoring devices.
`
`Peak counting? (not exactly
`specified)
`Statistics: ANOVA/t-test
`Manual counting of PPG
`peaks
`Band-pass filtering PPG sig-
`nal, then zero-crossing count.
`Take median of 10 consecu-
`tive HRs.
`Band-pass filtering PPG sig-
`nal, finding mean HR by au-
`tocorrelation analysis.
`
`a) Peak detection by
`threshold algorithm,
`reciprocal median pulse
`interval → mean HR.
`b) Looking for freq. with max.
`power in FFT spectrum.
`Analysis by PCC and
`mean ± SD of difference.
`
`Auto-detection method of
`monitoring devices. Test for
`|ΔHR| > 10 bpm
`
`Depending on condition,
`11.9–29% of HR values found
`unreliable (|ΔHR| > 10 bpm)
`
`High-pass filtering PPG sig-
`nal, then zero-crossing count.
`Evaluation of 30 s epochs and
`binary classification.
`Auto-detection method of
`monitoring devices.
`Bland-Altman analysis of
`ΔHR
`Manual elimination of
`motion artifacts. Visual
`analysis of PPG charts, man-
`ual counting of PPI. PCC of PR
`vs. HR
`Auto-detection method of
`monitoring devices.
`Bland–Altman analysis of
`ΔHR
`
`PPG recorded 1.1% (±0.7%)
`false negative and 0.9% (±0.6%)
`false positive beats, as
`compared to ECG.
`95% CI: ΔHR =−2 ±26 bpm
`
`Buttock: PCC = 0.999
`Leg: PCC = 0.995
`Back: PCC =−0.134 (sic!)
`
`95% CI: ΔHR =−0.4 ± 12 bpm
`
`Kornowski et al.
`(2003) [36]
`
`(Cardiac patients vs.
`healthy controls,
`ages 43 ± 18) N = 144
`
`Medic4All
`wristwatch
`telecare system
`
`Wrist (sensor
`worn like a
`watch)
`
`Vogel et al.
`(2007) [38]
`
`N = 1, T not specified,
`additionally 6 h of data
`from the “MIT-BIH
`normal sinus rhythm
`database”
`
`Self-developed
`IN-MONIT system,
`using both red
`and infrared LEDs
`
`Inside auditory
`canal
`
`Rest on chair,
`evaluation of up to 10
`intervals of 30 s for
`each patient,
`respectively.
`Not specified
`
`Neonates
`Barrington et al.
`(1988) [32]
`
`(Lung or cardiorespiratory
`disease) N=22×T=3 min
`
`Roche Medical
`Electronics 634
`
`“Gently
`restrained” limb
`
`Johansson et al.
`(1999) [33]
`
`N = 6 (4 preterm) ×
`T = 8 h
`
`Self-made PPG
`sensor, connected
`to HP 66S monitor
`
`Lateral side of
`left thigh
`
`Comparison of HR
`detected from PPG-
`and ECG-based
`monitors.
`Continuous
`monitoring during 8 h
`
`Kamlin et al.
`(2008) [39]
`
`N = 55 × T = 3 min
`
`Masimo Radical
`PO monitor
`
`Olsson et al.
`(2000) [41]
`
`N = 10
`(preterm) × T = 30 min
`
`Siemens photo
`diode SFH2030
`(940 nm)
`
`Right hand/wrist Experimenters
`reading HR displayed
`by PPG and ECG
`monitors from video.
`Rest in incubators
`
`Leg, buttock and
`interscapular back
`
`Singh et al.
`(2008) [40]
`
`N = 30 × T = 3 min
`
`Masimo Radical
`PO monitor
`
`Right hand
`
`Experimenters
`reading HR displayed
`by PPG and ECG
`monitors from video.
`
`values, in case b) they are HRV and PRV variables. In both approaches
`they are determined from detected cardiac cycles within the identical
`epoch of simultaneously recorded signals. In case c) the values under
`investigation consist of the time series of the detected cardiac cycle
`lengths themselves. Since each PPI belongs to an RRI, namely the
`one beginning directly before and separated only by PTT, one can
`also arrange them in pairs here. Whatever the case, one has a number
`of paired results derived from ECG and pulse signals.
`In order to assess the agreement of data pairs from two methods
`of measurement, the correct statistical approach is not obvious. Often
`researchers use Pearson's correlation coefficient (PCC); however, it
`quantifies linear correlation, not agreement. As examples to illustrate
`this, one can envisage situations where the second method either con-
`sistently overestimates values by an additive constant (location shift)
`or a multiplicative factor (scale shift) or both, which still yield a perfectly
`linear correlation. Aside from this, results of two techniques measuring
`the same data will almost always be linearly correlated. In addition,
`PCC tends to be greater, if the range of the true quantity in the sample
`is increased. Since investigators usually test a method in question
`
`with a wide range of data, a highly significant correlation is almost guar-
`anteed. The arguments speaking against PCC apply even more to Spear-
`man's rank correlation coefficient, which was used in one of the studies
`[27] summarized in this review.
`Testing the equality of the mean, e.g. by Student's t-test, is another
`insufficient way to verify agreement. It is merely able to detect
`whether the lack of accuracy of a method (i.e. its bias) is significantly
`worse than its lack of precision.
`In order to cure the shortcomings of the mentioned analyses,
`several approaches have been proposed. An example is the concor-
`dance correlation coefficient, a more suitable recast of the PCC [28].
`Most of the studies cited in this review, however, either simply
`itemize mean and standard deviation of the differences between
`ECG and pulse measures, or make use of the approach suggested
`by Bland and Altman [29]. In the latter, one plots the differences
`between the two methods versus the average as the best estimator
`of the true value. Even when one technique is considered as the
`gold standard, the average values should be plotted at the abscissa
`in order to avoid statistical artifacts [30]. A 95% confidence interval
`
`4
`
`

`

`A. Schäfer, J. Vagedes / International Journal of Cardiology 166 (2013) 15–29
`
`19
`
`500HzFinger
`200HzRightmiddlefinger
`
`500HzRightmiddlefinger
`
`Middlefinger
`
`Finapresdevice
`IHRasprovidedby
`Notspecified:used
`
`sensorinamattress
`b)Contactless
`1000Hza)Fingerclip
`
`b)Indexfinger
`1000Hza)Leftearlobe
`1000HzRightmiddlefinger
`
`400HzEarlobe
`
`shoulder,forehead
`forearm,wrist,
`
`66HzIndexfinger,medial
`
`100HzLeftearlobe
`
`400HzNotspecified
`
`Finger
`
`(reconstructedat2092Hz)
`128Hz
`
`200HzRadialartery
`1000HzRadialartery
`
`BiopacAcqKnowledgeII
`Self-madehandhelddevice
`
`1000HzNotspecified
`
`Colin-System
`
`OhmedaFinapres2300
`OhmedaFinapres2300
`
`OhmedaFinapres2300
`
`OhmedaFinapres2300
`
`→50×5min
`N=10,a)–e)5mineach
`
`InfraredPPGsensors(850nm)
`
`ADInstrumentsML305
`BIOPACtransducerTSD200
`
`1×5hofnightsleep
`15×4min
`15×6min
`
`ADInstrumentsPowerLab410
`
`44×6min
`
`aA/DconverterinaPC
`VariousPPGsensorslinkedto
`transmissionsensorPS-2105
`FeedbackInstrumentsPPG
`pulseoximetermodule
`NellcorMP506
`
`NellcorN-100pulseoximeter
`
`48×10min
`
`42×7min
`
`1724×2minepochsused
`Nightsleeprecords,
`
`40×30sanalyzed)
`(randomlyselected
`40×5min
`20×10s
`
`9×10min
`4minepochs
`subdividedinto
`Variousrecords,
`20×3×5min
`b)20×5min
`a)20×5min
`
`20HzDorsalsideofwrist
`
`30×6.0±0.8hofnightsleepDENSO“PrototypeC”
`
`Indexfinger
`
`(interpolatedto1000Hz)
`250Hz
`
`Leftmiddlefinger
`
`Exp.2)991Hz
`(downsampledto10–100Hz)
`Exp.1)1000Hz
`
`1000HzLeftfinger
`
`250HzIndexfinger
`
`TSD200transducer
`BiopacPPG100Camplifier+
`System1.5B
`2)FlexcompBiomonitoring
`
`1)BeckmanRMDynograph
`MP150+TSD123Btransducer
`BiopacPPG100Camplifier+
`sensor(860nm)
`BiopacMP30+SS4LPPG
`
`17×(4+5+4)min
`
`Exp.2)10×26min
`
`Exp.1)16×5min
`
`9×(10+20+20)min
`signals>1min)
`Notspecified(‘highquality’
`
`Positionofsensor
`
`Samplingrate
`
`Numberofrecords×durationDevice(s)used
`
`occlusion,followedbyreleaseperiod.
`Supinerest.Supradiastolicandsubsystolic
`Restinchair
`
`Treoetal.(2005)[49]
`N=40(hospitaloutpatients)
`Kristiansenetal.(2005)[48]N=20(healthy,ages19–51)
`Impedanceplethysmography(IP)studies
`
`Supinerest
`
`cardiomyopathy,ages45.1±7.3)
`N=9(patientswithdilated
`
`Suhrbieretal.(2006)[47]
`
`physicalormentaltasks
`Supineorsittingrest,various
`Supinerest
`4mincontrolledbreathing
`a)supineb)orthostatic1minspontaneous,
`
`studies,variousageranges)
`N=234(patientsfrommultiple
`N=20(healthy,ages22–75)
`and10controls,bothaged6–13)
`N=20(10childrenwithpace-maker
`
`McKinleyetal.(2003)[46]
`Dawsonetal.(1998)[44]
`
`Constantetal.(1999)[43]
`
`c)standingd)exercisee)recovery
`a)supineb)supineandcontrolledbreathing
`
`N=10(healthy,ages22.5±1.6)
`
`Carrascoetal.(1998)[42]
`Continuousbloodpressure(CBP)studies
`
`Supinerest
`
`N=15(healthy,ages23–38)
`
`Wongetal.(2010)[65]
`
`14×10min
`10×5min
`
`Supinerest
`Rest(nototherwisespecified)
`andcontrolledbreathing,respectively
`Restinchair;3minofspontaneous
`
`N=14(healthy,ages25.8±4.2)
`N=10(healthy,ages21–28)
`
`Shietal.(2008)[54]
`Selvarajetal.(2008)[53]
`
`N=44(healthy,ages16–60)
`
`Rauhetal.(2004)[27]
`
`Lyingsemi-erectinhospitalbed
`
`N=48(healthy,ages20–68)
`
`Nilssonetal.(2007)[67]
`
`Semi-recumbentrest
`
`N=42(healthy,age21±3)
`
`Luetal.(2009a)[58]
`
`10mininuprightpos.,10mininsupinepos.10×(10+10)min
`
`Nightsleep
`pulsedetectionbydevice)
`Nightsleep(otherwisenostable
`
`N=10(healthy,ages26±7.5)
`(OSA,ages54.9±16.3)
`
`Luetal.(2008)[57]
`
`Khandokeretal.(2010)[105]N=29(healthy,ages51.3±8.5)+22
`
`N=30(healthy,ages22–47)
`
`Hayanoetal.(2005)[55]
`
`Tilttable:supine–head-uptilt–supine
`
`N=17(healthy,ages28.5±2.8)
`
`Giletal.(2010)[56]
`
`Exp.2)mentalexercisetests
`
`Exp.2)N=10(healthy,ages25–50)
`
`Exp.1)restinchair
`slowwalk,bicycleexerciseinsupinepos.
`3sympatheticstimulations:orthostatictest,
`
`Exp.1)N=16(healthy,ages23–35)
`
`Giardinoetal.(2002)[52]
`
`N=9(healthymen,ages26.3±4.5)
`
`Charlotetal.(2009)[59]
`
`Restinchair
`
`N=10(healthymen,ages25–35)
`
`Changetal.(2007)[104]
`Photoplethysmographicstudies
`
`Experimentalpostureorconditions
`
`Nandtypeofsubjects
`
`Study
`
`SurveyofstudiesontheaccuracyofPRVcomparedtoHRV.Foranalysismethodsandresultscf.Table5.
`Table4
`
`5
`
`

`

`20
`
`A. Schäfer, J. Vagedes / International Journal of Cardiology 166 (2013) 15–29
`
`distributions
`PPIandRRI
`kurtosissimilarfor
`Skewnessand
`inconspicuous(t0)
`variables
`t-testsforall
`
`pacemakersubjects
`fornormaland
`supineandstanding,
`Resultsvalidfor
`respectively
`(almostall)HRVvars,
`differenceforall
`findsignificant
`(duringstanding)
`t-testsduringexercise
`
`ingrest
`HF(logarithms)dur-
`BA+forLF,BA−for
`
`studies
`datafromearlier
`secondaryanalysesof
`Resultsobtainedby
`
`inconspicuous(t0)
`variables
`t-testsforall
`
`overestimateHRV
`mostlytendsto
`available:PRV
`t-testsalso
`
`TP:BA+(c++)
`
`BA+(c++)
`
`BA+(c+)
`
`BA+(c++)BA+(c++)BA+(c++)
`
`ApEn:(c0)
`ApEn:(c0)
`
`(c++)
`
`c−
`
`BA+(c+)
`
`BA+(c++)
`
`t−
`
`(t0)
`
`TP:(c++)Δ+
`TP:c−Δ−
`TP:(c+)Δ−
`TP:(c++)Δ+
`TP:(c++)Δ++
`
`(c0)Δ0
`c−Δ−
`(c0)Δ−
`(c+)Δ+
`(c++)Δ+
`
`(c+)Δ0
`c−Δ−
`(c0)Δ−
`(c++)Δ0
`(c++)Δ+
`
`TP:BA−(c++)
`VLF:BA0(c++)
`
`BA−(c+)
`
`(c++)t−
`
`c−t−
`
`BA−(c++)BA−(c++)
`
`BA−(c+)
`
`(c++)t−
`
`c−t−
`
`(c+)t−
`(c++)t−
`
`SD1,SD2:BA+
`
`BA+
`
`BA+
`
`SD2:(c++)
`SD1:(c0)
`
`SampEn:u−
`SampEn:u−
`
`(c0)
`
`u−
`(u0)
`
`u−
`(u0)
`
`BA−
`
`BA+
`
`(c0)/c−
`
`(c0)
`(c0)
`
`(c+)
`
`(c+)
`(c++)
`(c++)
`
`(c++)
`
`BA0
`
`c−
`
`BA0
`
`c−
`
`(c0)
`
`(c0)
`
`u−
`(u0)
`
`(c0)
`
`(c++)
`
`(c+)
`(c++)
`
`(c+)
`(c++)
`
`(c++)
`(c++)
`
`(c++)
`(c++)
`
`BA−(c++)
`differs(−)
`Statistics
`
`BA−c−
`BA−c−
`BA−c−
`
`BA+(c++)
`
`BA−c−
`BA−c−
`BA−c−
`
`BA0(c++)
`BA0(c++)
`
`BA−c−
`BA−c0
`
`BA−(c+)
`BA−(c++)
`BA0(c++)
`BA+(c++)
`
`LF/HFratio
`
`[n.u./%]
`LFandHF
`
`HFpower
`
`LFpower
`
`pNN50
`
`Remarks
`
`Othervariables
`
`HRVfrequencydomainvariables
`
`BA+(c+)
`
`BA+(c++)
`
`BA++(c++)
`
`(c+)
`(c++)
`
`(c++)
`(c++)
`
`oftotalpower)
`(Diff.=1.35–3%
`
`Standing
`Supine
`
`Shi2008[54]
`
`Lu2008[57]
`
`BA0(c++)
`
`BA++(c++)
`
`BA++,(c++)
`
`Kristiansen2005[48](IP)
`Hayano2005[55]
`
`Detectionofdiastolicminima.
`manufacturer'ssoftware.
`andarrhythmiaswithdevice
`Preprocessingforartifacts
`negativeslopeincatacroticphase.
`(EMD).Detectingmaximum
`Empiricalmodedecomposition
`samplepulsecycle
`Crosscorrelationwitha
`Pulsefr