Introduction

Blood pressure variability (BPV) has emerged as a hemodynamic marker of interest with widespread clinical and research utility. BPV is associated with increased mortality1, and is a risk factor for cardiovascular disease2,3,4,5,6,7,8,9, stroke10,11,12, cerebral small vessel disease (CSVD)13,14, neurodegenerative disease15 and dementia16,17, independent of average blood pressure (BP). BPV appears to be of particular importance in older adults while mean BP might be more important in middle aged adults18.The regulatory mechanisms which underlie BPV modulation are complex and multifactorial, potentially involving arterial baroreflex sensitivity19, arterial stiffness20, endothelial dysfunction21, and kidney function21 among other factors22,23. BPV has been assessed in multiple ways, but little is known about the reliability of any one type of assessment or the differences in reliability between assessments.

Methodologies for BPV calculation fall into four main categories including very short-term (beat-to-beat) BPV, short-term BPV (< 24 h), medium-term BPV (day-to-day), and long-term BPV (visit-to-visit over months or years)2. Each category offers unique advantages and disadvantages. Long-term and medium-term BPV are influenced primarily by environmental and behavioral factors such as season, altitude, and antihypertensive medication adherence24,25,26. Short-term BPV is reflective of circadian BP rhythms like nocturnal BP dipping and morning BP surge26,27 and is influenced by central and peripheral autonomic modulation and arterial elasticity25,28,29,30. Very short-term beat-to-beat BPV dynamics yield detailed insights into autonomic and cardiovascular function24,31,32, but have previously required invasive arterial catheterization. The validation of noninvasive continuous arterial BP measurement technology33,34,35,36,37 now enables an accessible and noninvasive methodology for generating validated continuous BP waveforms without arterial catheterization. Numerous hemodynamic markers of interest can be derived from the raw arterial pressure waveforms generated, including standard deviation (SD), coefficient of variation (CV), variability independent of the mean (VIM), and average real variability (ARV) (see methods). The availability of MRI-compatible devices allows for the study of beat-to-beat BP dynamics during neuroimaging, which may be relevant to cerebrovascular and neurodegenerative conditions associated with increased BPV38,39,40,41.

Numerous studies have identified relationships between BPV metrics and clinically relevant markers of pathophysiology, but no systematic investigation has yet been performed assessing reliability and delineating specific use cases for each metric. Reporting of specific BPV metrics is therefore largely driven by outcomes or convenience. A first step towards elucidating the specific utility and rationale for using a particular BPV metric is to establish its test–retest reliability. However, despite the demonstrated utility of noninvasive continuous BP monitoring, little is known about the reliability of beat-to-beat BPV metrics. The ability to measure BPV accurately and reliably has also been questioned by some42. Few studies have assessed BPV test–retest reliability, of those that have, most investigations focusing on visit-to-visit BPV43,44. To the authors’ knowledge, only one study has assessed the test–retest reliability of any beat-to-beat BPV metric, but this was performed in the context of orthostatic beat-to-beat blood pressure response45.

The present study addresses this knowledge gap by examining the intrasession test–retest reliability of beat-to-beat BPV metrics in a sample of community dwelling older adults recruited as part of the vascular senescence and cognition (VaSC) cohort. As part of the VaSC study, beat-to-beat BPV was monitored during brain MRI using an MRI-compatible noninvasive continuous BP device. These data are leveraged in the present study to examine reliability of BPV metrics, including beat-to-beat systolic and diastolic SD, CV, VIM, and ARV, as well as SBPrange (maximum SBP minus minimum SBP). The test–retest reliability of noninvasive continuous average systolic BP (SBP), diastolic BP (DBP), and heart rate (HR) are also assessed.

Results

121 participant visits with continuous BP monitoring were available. Of these, 10 visits were excluded due to poor data quality and excessive motion artifacts. After exclusion, 111 participant visits were included for analysis. Participant characteristics and demographics for this sample are displayed in (Table 1). The correlation between all BPV metrics and average blood pressure with and without demographic adjustment, as well as the correlation between all BPV metrics and age is shown in (Supplementary File 1).

Table 1 Participant characteristics and demographics (N = 111).
Full size table

Blood pressure variability intrasession test–retest reliability

One Table 2 shows the sample mean for each analyzed cardiovascular parameter as well as the paired t-test p-value associated with each test–retest comparison. No significant differences between waveform 1 (test) and waveform 2 (re-test) were observed for any measure. Violin plots display waveform 1 and waveform 2 data distributions for selected measures in (Fig. 1).

Table 2 Cardiovascular parameters derived from two sequential 7 min continuous blood pressure waveforms.
Full size table
Fig. 1
figure 1

Test–retest reliability of beat-to-beat blood pressure variability metrics. Test–retest violin plots with paired t-test comparison (row 1), test–retest correlation plots with intraclass correlation coefficients (ICC) (row 2) and Bland–Altman plots with bias and 95% limits of agreement indicated by horizontal dashed lines shown for five measures of systolic blood pressure variability: (A) Systolic blood pressure average real variability (SBP ARV), (B) Systolic blood pressure standard deviation (SBP SD), (C) Systolic blood pressure coefficient of variation (SBP CV, (D) Systolic blood pressure variability independent of the mean (SBP VIM), and (E) maximum minus minimum systolic blood pressure (SBPrange).

Full size image

All noninvasive BP waveform-derived cardiovascular parameters displayed excellent absolute test–retest reliability (SEM < 5%, SRD < 10%). HR, SBP, and DBP means displayed excellent relative test–retest reliability (ICC > 0.90), all dispersion (SD, CV, VIM) and sequence (ARV) BPV metrics displayed good relative test–retest reliability (ICC = 0.70–0.89), while SBPrange displayed moderate relative test–retest reliability (ICC = 0.50–0.69). These results are shown in (Table 3). Correlation and Bland–Altman46 plots are displayed in (Fig. 1).

Table 3 Test–retest reliability of cardiovascular parameters obtained from continuous noninvasive blood pressure monitor.
Full size table

Antihypertensive treatment and BPV test–retest reliability

SBP ARV displayed excellent test–retest reliability in the group taking no antihypertensive medications, and good test–retest reliability in the group taking antihypertensive medications. All systolic BPV dispersion measures displayed good reliability in the no hypertensives group, but only moderate reliability in the antihypertensive group. Similar observations were seen for the diastolic BPV metrics, except for DBP VIM, which showed improved reliability in the antihypertensive group. Results displayed in (Table 4).

Table 4 Test–retest reliability of blood pressure variability metrics obtained from continuous noninvasive blood pressure monitor stratified by antihypertensive medication use.
Full size table

Discussion

The present study finds that sequence and dispersion-based measures of beat-to-beat systolic and diastolic BPV can be reliably derived from a continuous BP monitoring device. Sequence-based metrics, including systolic and diastolic ARV, displayed the highest test–retest reliability in the overall sample. This is likely due to their decreased susceptibility to outliers and low-frequency oscillations in beat-to-beat BP compared to dispersion and instability BPV metrics47,48,49. This concept is illustrated in (Fig. 2A,B), where we can see BP dispersion metrics, such as SD, are more heavily influenced by low-frequency oscillations in BP, while BP ARV is more directly influenced by beat-to-beat changes in blood pressure.

Fig. 2
figure 2

Visualization of two distinct blood pressure variability profiles (A) An example waveform from an individual with high blood pressure standard deviation (BP SD) but low average real variability (ARV). (B) An example waveform from an individual with high ARV but low BP SD.

Full size image

Low-frequency oscillations may be modulated in part by changes to peripheral vascular resistance50, transient oscillatory responses to hemodynamic perturbations51, and intrinsic vasomotor rhythmicity50 while beat-to-beat changes in BP are mediated by central sympathetic drive, arterial and cardiopulmonary reflexes, and arterial stiffness26,52. Additionally, ARV considers the temporal order of BP measurements, adding a time series variability component to the measurement53 since it reflects the variation in successive differences in beat-to-beat BP. These features of ARV potentially add prognostic value54 and overcome some pitfalls of the SD-based measures which only measure dispersion around mean BP and may be more influenced by outliers55, while also ignoring the temporal order of BP measurements2,54,56,57,58. Additionally, two individuals with different BP profiles may have similar BPV dispersion measures but different ARVs59. The fact that ARV entails measurement of consecutive beat-to-beat differences rather than dispersion from the mean make ARV particularly well-suited for assessing beat-to-beat variation in BP2. A parallel can be drawn here to a heart rate variability (HRV) metric, root mean square of successive differences (RMSSD)60, which acts as a high pass filter thus reflecting the high frequency variability in heart rate and is calculated similarly to ARV60. RMSSD has been shown to offer certain advantages over other HRV metrics, and it’s possible that ARV may share some of these same advantages such as shorter required sampling durations for reliable measurements61, and unique insights into parasympathetic tone62, but further research is needed.

Conversely, SBPrange, a measure of systolic BP instability, displayed the lowest test–retest reliability, likely due to increased susceptibility to outliers and swings in BP over time47. All six systolic and diastolic BP dispersion metrics displayed good test–retest reliability. The similar reliability between dispersion measures means that metric choice should be based on context and the individual characteristics of each measure.

When stratified by antihypertensive medication use, SBP ARV displayed excellent reliability in the group not using antihypertensive medications and was the only systolic BPV metric that displayed good reliability in the group using antihypertensive medications. All other measures of systolic BPV displayed moderate or poor test–retest reliability. This is a clinically relevant finding given the widespread use of antihypertensive medication in older adults63, and supports the use of SBP ARV regardless of antihypertensive treatment status.

The mechanisms responsible for increased systolic BPV are more clearly understood than those associated with increased diastolic BPV. For example, visit-to-visit systolic BPV has been shown to correlate with arterial stiffness64 and worsening renal function65, while visit-to-visit DBP variability has not. A key difference between visit-to-visit BPV and beat-to-beat BPV is that visit-to-visit BPV could be influenced by antihypertensive medication adherence66. For this reason, beat-to-beat BPV may be a more accurate assessment of the underlying physiology which modulates BPV. Regarding this underlying physiology, although not fully understood55, most studies place central importance on the central sympathetic drive24,26, neuronal reflexes22,24,67,68, and arterial stiffness64,69,70.

The variety of available BPV metrics and relative lack of understanding pertaining to causal mechanisms has resulted in metric choice being largely based on convenience up to this point, and most studies have used less granular and likely less reliable visit-to-visit measures. More investigations should be conducted to differentiate the underlying mechanisms that modulate beat-to-beat BPV metrics and continued assessment of reliability is needed. Limitations of the current study include a relatively small sample size and a restricted age range potentially limiting generalizability.

The present study supports the use of continuous noninvasive BP derived metrics of BPV in older adults. Of those metrics tested, ARV displayed the highest level of test–retest reliability, perhaps due to its decreased susceptibility to outliers and low frequency oscillations in BP. While these low frequency oscillations reduced the reliability of beat-to-beat dispersion BPV measures, they may not be entirely extraneous, and should therefore continue to be studied to fully capture the multi-dimensional nature of BPV. All BPV metrics displayed good or excellent test–retest reliability in the present investigation, except for maximum minus minimum systolic BP. Future studies investigating the effects of beat-to-beat BPV should include ARV due to its increased reliability regardless of antihypertensive treatment status, and sensitivity to consecutive beat-to-beat differences in BP.

Methods

Participants

Participants were recruited from Los Angeles County and Orange County communities, and all procedures were conducted as part of the VaSC Study at the University of Southern California (USC) and University of California Irvine (UCI). Older adults aged 55 to 89 years who were living independently were included. Exclusion criteria were history of clinical stroke, dementia, major neurological or psychiatric disorder or medications impairing the central nervous system, current organ failure or other uncontrolled systemic illness, or contraindication for brain MRI. Study inclusions and exclusions were verified by a structured clinical health interview and review of current medications with the participant and, when available, an informed study partner. This study was approved by the University of Southern California (HS-14-00784) and University of California, Irvine (HS-2019-5324) Institutional Review Boards, all participants gave informed consent, and the study was performed in accordance with all relevant guidelines and regulations.

Continuous BP data processing and analysis

Participants were asked to take medications as normally prescribed and abstain from caffeine the morning of data collection. Beat-to-beat BP measurements were obtained continuously during supine rest in a 3 T Siemens MRI scanner, using an MRI compatible non-invasive continuous BP finger cuff device (Biopac®). First, the participant rests for 3 min in the supine position prior to the calibration period. During calibration, BP waveforms are acquired by the continuous monitoring device and 2 static pressures are simultaneously acquired using a calibrated, MRI compatible automatic BP device with an inflatable brachial artery cuff (TeslaDUO). These static pressures are used to calibrate the continuous BP monitor using the Caretaker® system (Biopac®). After calibration, continuous BP was monitored during 2 sequential, 7 min MRI scans.

The Calib upsample utility (Biopac®) was used to extract continuous arterial pressure data obtained during the 2 sequential, 7 min MRI scans at a sample rate of 100 Hz. Data segments free from obvious motion artifacts were selected from each 7-min continuous BP data segment for further processing. Waveforms were excluded if more than 10% of the data needed to be excluded to remove obvious motion artifacts. Two 420 s waveform segment examples, one with an obvious motion artifact and one without, are shown in (Fig. 3) for illustration purposes.

Fig. 3
figure 3

Illustration of visual inspection for motion artifacts. (A) 420 s continuous blood pressure (BP) waveform with accurate peak (systolic blood pressure) and trough (diastolic blood pressure) detection free from obvious motion artifacts. (B) 420 s continuous BP waveform with motion artifact visible at approximately 2693 s.

Full size image

A peak detection algorithm was used to identify SBP peaks which served as the basis for further cardiovascular parameter calculation. Peaks were detected using the find_peaks function from the scipy.signal library71, with default parameters set to a minimum detection height of 80 mmHg, and a minimum peak separation of 400 ms by default. Diastolic troughs were identified as the lowest BP reading between two systolic peaks. Each waveform was then visually inspected using the VaSC BP Signal Toolbox application for erroneous or missing peaks and troughs by TL. Occasionally, default data filtering parameters were adjusted as needed to ensure accurate peak and trough detection. A visual illustration of this process is shown in (Fig. 4A,B).

Fig. 4
figure 4

Illustration of peak detection parameter adjustment. (A) A 10 s continuous blood pressure waveform with systolic peak and diastolic trough detection applied. Two pronounced dicrotic notches are erroneously detected as peaks by the peak detection algorithm at the 200.5 and 201.6 s marks when using the default minimum detection height and distance parameters (80 mmHg and 400 ms). (B) The same waveform displayed in (Fig. 2A) is now displayed with a modified minimum distance peak detection parameter of 450 ms. This modification results in accurate peak detection across the waveform.

Full size image

Calculation of blood pressure variability metrics

In addition to measurement time, BPV metrics can also be categorized by index type (frequency, dispersion, sequence, or instability)55,72. Three measures of BP dispersion (SD, CV, and VIM), 1 measure of BP instability (SBPrange), and 1 measure of BP sequence (ARV) were calculated for test–retest comparison in the present study.

The standard deviation of SBP and DBP amplitude measurements (Fig. 5A) was obtained across the waveform’s duration as shown in (Fig. 5B,C). BP SD was then further processed into CV and VIM3. Similar processes were repeated for diastolic BP metrics. Of these metrics, BP SD is used most often due to its straightforward calculation and interpretation, however, it may be correlated with sample mean BP26,73. BP CV and BP VIM measures may compliment BP SD because they are generally reported as less strongly correlated with mean BP11,25,42,74, allowing for comparison of samples with different means in the case of CV75 without average BP adjustment. BP CV is calculated as (BP SD/BP mean)*100 (Fig. 5B), while BP VIM is calculated by taking BP SD readings divided by mean BP raised to the power of x (SBP VIM x = 0.46), where x was derived from a non-linear curve fitting of BP standard deviation (SD) against average BP using the nls package in R76 (Fig. 5B).

Fig. 5
figure 5

Calculation of blood pressure variability metrics. (A) A single arterial pulse waveform with systolic blood pressure (SBP) peak and diastolic blood pressure (DBP) trough indicated. (B) Five blood pressure variability metric formulas used in the present analysis: Blood pressure (BP) standard deviation, coefficient of variation (CV), variability independent of the mean (VIM), SBPrange, and average real variability (ARV). (C) A visual representation of continuous BP waveform with 5 arterial pulse cycles and SBP measurements used for systolic blood pressure variability calculations.

Full size image

The difference between the maximum SBP reading and minimum SBP reading (SBPrange) was included as a measure of BP instability. SBPrange is the difference between the maximum and minimum systolic BP readings in a specified window, 7 min for the present study.

Systolic and diastolic ARV measures were calculated by taking the absolute differences between consecutive peaks and troughs respectively, and then averaging them across the 7 min continuous BP waveform26,56 (Fig. 5B). To further confirm the reliability of the continuous BP monitoring methodology the intrasession test–retest reliability of HR and BP were also assessed by comparing the mean HR, mean SBP, and mean DBP across each selected 7 min waveform.

Data analysis

All statistical analyses were carried out using R76. Paired t-tests were used to compare mean values of waveform 1 and 2 (test–retest). Intraclass correlation coefficient (ICC) with a 95% confidence interval was used to assess relative reliability using Munro’s criteria77 for interpretation while absolute reliability was assessed using the standard error of measurement (SEM), SEM%, smallest real difference (SRD), and SRD%. SEM is calculated as the SD of differences between paired measurements divided by the square root of the sample size78,79 while SRD represents the smallest change in a measurement that likely represents a true change rather than a measurement error78,80. ICC, SEM, and SRD are commonly used measures of test–retest reliability and are specifically used for this purpose in literature81,82,83,84.