Single-source PPG-based local pulse wave velocity measurement: a potential cuffless blood pressure estimation technique

The arterial blood pulse signal acquired using PPG transducers consisted of both static (DC) and pulsating (AC) components (Kamal et al 1989, Lindberg et al 1992). The DC component of the PPG signal was proportional to the total blood volume in the measurement area. The AC component was synchronized with the cardiac cycle and followed the arterial blood volume and flow changes. The AC component of the PPG signal was typically less than 0.2% of the DC amplitude. Removal of DC from the PPG waveforms was mandatory for the amplification of the AC components without causing permanent saturation of the amplifiers. In practice, it not straightforward to remove DC components from the PPG signals without introducing any additional signal propagation delay. Direct use of resistor–capacitor-based signal-conditioning filters for DC elimination causes significant inter-channel delay due to the mismatch between their characteristic time constants. A non-zero inter-channel delay was critical for transit time delay measurements employing sensors placed in close proximity. Hence, an AFE circuit with negligible inter-channel delay and a high-speed signal acquisition module were desired for reliable blood pulse detection.

An application-specific dual-channel AFE was designed and developed for simultaneous acquisition of PPG signals. The basic circuit schematic of the AFE, internal circuit diagram of single-source PPG probe and software architecture of the proposed measurement system is described in figure 2. The proposed AFE design was primarily focused on reducing the inter-channel delay between analog channels. Extraction of pulsatile components from the raw blood pulse signals using DC subtraction (without any signal-conditioning filters and energy storage elements in the principal signal path) significantly reduces the inter-channel delay (Nabeel et al 2017). As illustrated in figure 2, the output current signals from the photodiodes of the PPG transducers were initially converted into a voltage signal using transimpedance amplifiers. The average DC voltage of the PPG signal was extracted using an R–C network (corner frequency  =  0.05 Hz) implemented in the parallel signal path. The raw PPG signal and its average DC voltage level from each channel were given to the $V_{{\rm IN}}^{+}$ and $V_{{\rm IN}}^{-}$ input terminals of precision instrumentation amplifiers (Texas Instruments—INA2126), respectively, via Opamp (Texas Instruments—TL082CP) based voltage buffers. Subtraction of the DC component from the original PPG signal and further amplification (gain  ≈  40 dB) were performed using an instrumentation amplifier. Amplified PPG signals were digitized using a simultaneous sampling data acquisition (DAQ) module (National Instruments^®—PXIe-6368). Signals were continuously acquired at a sampling rate of 100 kS/s/channel with 16-bit resolution. Use of a simultaneous sampling DAQ card with high sampling rate helped to eliminate any mismatch due to digital pipelining, data transfer delays and timing error arising from the system sampling clock.

Zoom In Zoom Out Reset image size

Signal acquisition and processing were performed using a custom program developed using LabVIEW (National Instruments^®). Digital domain processing, transit time measurement, and PWV calculation were automatically performed. The measurement algorithm (figure 2) was divided into two steps: signal processing and pulse wave analysis. In the signal processing stage, acquired PPG waveforms were low-pass filtered using a Butterworth second order digital filter. Since the expected range of blood pulse signal frequency in humans is from 0.7 Hz to 3 Hz, the low-pass filter was realized with 10 Hz cut-off frequency, thus allowing a fundamental and at least two harmonics, while eliminating the out-of-band noise. The zero-phase difference between the raw signal and the filtered signal was ensured by applying a filter operation in the forward and backward directions (zero-phase shift filter operation). Further, the baseline drift of the PPG waveforms caused by body movement and respiration was removed using a wandering removal process (Fedotov et al 2014). Finally, cycle cutting was performed to separate each beat from the signal train and to collate the corresponding proximal and distal cardiac cycle pairs. For this, the diastolic minimum points of the proximal PPG signal were automatically detected from each beat. 50 ms prior to the diastolic minimum was taken as the fiducial point for cycle cutting. The time axis of the proximal and distal cardiac cycle pairs was synchronized by taking the same fiducial points in both channels. These cycle pairs were used for beat-by-beat local PWV measurement. The LabVIEW program also provided provision for real-time signal storage of acquired waveforms for future reference.

The arterial pulse signal is composed of a forward moving blood pulse wave superimposed with reflected signals from various bifurcation points in the arterial tree (Boutouyrie et al 2009). The reflected signals travel with different propagation velocity and amplitude. Depending on the arterial sites, the reflected signals alter the shape and timing information of original pulse signal. Since the early systolic region of blood pulse is minimally affected by the reflected waves (Hermeling et al 2007), the characteristic points for transit time measurements are often selected from this region. Characteristic points are specific identification points based on the physiological features of the blood pulse waveform. The time delay between characteristic points of the proximal and distal cycle pair is referred to as transit time (ΔT). Therefore, precise identification of characteristic points is highly desirable for accurate ΔT and PWV calculation.

3.3.1. Characteristics point identification.

The second derivative maximum technique (Kazanavicius et al 2005) was implemented for characteristic point identification. It was reported as an efficient method for repeatable measurements (Chiu et al 1991). The second derivative maximum technique uses the point at which the second derivative of a blood pulse cycle is maximal as its characteristic point. It is typically close to the systolic foot position, where the change in slope is a maximum. For a given cardiac cycle with dataset values (Ψ) collected at time intervals (t) having a finite sample number (k), the second derivative can be determined using the formula given in (1).

$$\begin{align} \newcommand{\e}{{\rm e}} \displaystyle \frac{{{{\rm d}}^{2}}\Psi }{{\rm d}{{t}^{2}}}=~\frac{\left( {{\Psi }_{k+1}}-{{\Psi }_{k}} \right)-({{\Psi }_{k}}-{{\Psi }_{k-1}})}{\frac{1}{2}({{t}_{k+1}}-{{t}_{k-1}})}\nonumber \end{align} \tag{ 1 }$$

3.3.2. Local ΔT and PWV measurement.

The pulse wave analysis phase emphasizes accurate measurement of ΔT to obtain reliable carotid local PWV values. Using a characteristic point detection algorithm, the local ΔT was calculated from the fiducial points of the proximal and distal carotid pulse cycle pair (figure 3). The centre-to-centre distance between the photodiodes was taken as the blood pulse propagation distance (ΔD). Within a small distance (ΔD  =  24 mm), the carotid artery was assumed to be a homogeneous section with no curvature. Finally, the carotid local PWV was estimated from each cardiac cycle in a beat-by-beat manner using (2).

$$\begin{align} \newcommand{\e}{{\rm e}} \displaystyle {\rm local}\,{\rm PWV}=~\frac{\Delta D}{\Delta T}\nonumber \end{align} \tag{ 2 }$$

Zoom In Zoom Out Reset image size

Real-time finger PPG and ECG signals were also acquired along with dual carotid PPG for simultaneous measurement of local PWV, PTT_R, and PAT_C. Synchronized acquisition of all four waveforms was performed using the developed local PWV measurement system advanced with two sub-units: one for the finger PPG signal acquisition and another one for real-time ECG signal acquisition. The carotid PPG AFE design was replicated in the finger PPG channel for signal acquisition using a transmittance-type fingertip PPG probe. The reference ECG signal was acquired using a commercial three-electrode system with a fully integrated single-lead AFE (Analog Devices Inc.—AD8232). Signals were digitized using a simultaneous sampling DAQ card (National Instruments^®—PXIe-6368) at a sampling rate of 100 kS/s/channel. The signal processing software was also modified for simultaneous measurement of local PWV, PTT_R, and PAT_C. As illustrated in figure 3, PTT_R was determined as the time delay between the fiducial points of the proximal carotid PPG signal and the finger PPG signal. PAT_C was determined as the time delay between the R peak of the QRS complex of the ECG and the fiducial point of the proximal carotid PPG waveform. Figure 4 shows the developed prototype of a single-source PPG-based local PWV device advanced with ECG and finger PPG signal acquisition modules.

Zoom In Zoom Out Reset image size

• To verify the efficiency of the single-source PPG probe and associated measurement system for continuous signal acquisition, repeatable and reproducible local PWV measurement from multiple subjects.
• To perform an inter-subject correlation analysis of carotid local PWV, PTT_R, and PAT_C versus brachial BP parameters.
• To examine the proposed local PWV device regarding its ability to measure the beat-by-beat variation of local PWV due to BP changes in individual subjects.
• To test the cuffless BP estimation accuracy of the proposed system with inter- and intra-subject models after best-case calibration.

Thirty-five normotensive subjects (age  =  28  ±  4.5 years; gender: 23 males and 12 females; height  =  168  ±  12 cm; weight  =  73  ±  15 kg) were recruited for the in vivo validation study. Written informed consent was obtained from all participants after explaining the study objectives and protocols. Anthropometric measurements were initially taken and recorded in the subject database. Participants were requested to take 5 min rest in a temperature controlled room (23  ±  2 °C) before the study to achieve a steady heartrate and BP state.

5.3.1. Blood pulse signal acquisition and local PWV measurement.

5.3.2. Local PWV reproducibility evaluation.

5.3.3. BP measurement from the brachial artery.

5.3.4. Simultaneous measurement of local PWV, PTT_R, and PAT_C.

5.3.5. Local PWV measurement during post-exercise relaxation.

Before the detailed analysis of in vivo study observations and results, a brief discussion about the accuracy of the measurement system and potential error sources in the recorded carotid local PWV are presented. Ambiguity in ΔD was controlled by precision placement of photodiodes during the final transducer assembly process. ΔD (=24 mm) was measured using a digital Vernier caliper (Mitutoyo—Digimatic Caliper—CD-6''VC) with an instrumental error of  ±0.03 mm. Hence, a maximum of 0.125% error in ΔD (%ε_ΔD  =  0.125%) was expected in the prototype design of the single-source PPG probe.

Hardware inter-channel delay and the system sampling clock can introduce timing errors in local PWV estimates. The inter-channel delay obtained from five selected subjects with different heartrate and BP is presented in figure 6. It was measured by feeding the carotid PPG signal from a single sensor to both the input channels simultaneously. The band inside each box in figure 6 indicates the median of the inter-channel delay values obtained from multiple cycles. The bottom and top of the boxes indicate the first and third quartiles with vertical lines extending to the minimum and the maximum delay values, respectively. A maximum of 1.8% error in local ΔT (%ε_ΔT  =  1.8%) was expected due to hardware inter-channel delay. Timing errors caused by the sampling of the acquisition module (sampling rate  =  100 kS/s/channel; clock period  =  10 µs) was insignificant when compared to the expected local ΔT values. Further, the tissue transit time of infrared light waves is much shorter than the blood PTT. Any difference in tissue transit time at two measurement locations can be neglected due to similar skin and tissue composition within a small section. Therefore, the expected error in the carotid local PWV (%ε_{local PWV}  =  %ε_ΔD  +  %ε_ΔT) due to hardware limitations was less than 2%.

Zoom In Zoom Out Reset image size

Proximal and distal blood pulse signals for local PWV measurement were acquired from multiple subjects under different physiological conditions. The measurement algorithm was susceptible to the quality of the acquired blood pulse waveforms. The prototype device was able to acquire dual PPG waveforms with a signal-to-noise ratio of approximately 28 dB, which was adequate for reliable local ΔT measurement using the second derivative maximum technique (Solà et al 2009). Figure 7 exhibits carotid blood pulse waveforms (after signal conditioning) acquired from a volunteer (age  =  26 years, body mass index (BMI)  =  25.05 kg m⁻², SBP/DBP  =  118/73 mmHg). The competency of the single-source PPG transducer to acquire distinct pulse waveforms from adjacent sites makes the proposed design perfect for local PWV measurement. The peak-to-peak amplitude of raw pulse waveforms was subjective and varied between 6 to 17 mV_PP in individual subjects during the present study. The inter-beat period of the acquired PPG signals was compared against R–R interval of ECG waveform to ensure reliability. The mean absolute error in the inter-beat period measurement was less than 0.14%.

Zoom In Zoom Out Reset image size

The prototype design of a single-source PPG probe was efficient and facilitates easy measurement of the beat-by-beat carotid local PWV. Artery detection, probe placement and orientation procedures, signal acquisition, and measurement process were completed within 2 min. Baseline carotid PWV estimates were less compared to the previously reported carotid-to-femoral and carotid-to-radial PWV (Reusz et al 2010). Mean carotid local PWV values obtained from 35 subjects were in the range 2.52  ±  0.41 m s⁻¹. This was consistently observed for multiple subjects with high reproducibility under various physiological conditions. The obtained values were comparable with the carotid PWV reported in previous studies using ultrasound and accelerometer transducers (Sorensen et al 2008, Faita et al 2010). Figures 8(a)–(c) show the recorded beat-by-beat carotid local PWV values of three different volunteers (30 s data). The beat-to-beat variation was calculated by dividing the SD by the mean and expressing it as a percentage. The minimum and maximum beat-to-beat variation observed during the study were 4.15% and 11.38%, respectively. Relatively high beat-to-beat variation (7.01–11.38%) was observed in subjects with obesity and weak carotid surface pulsations.

Zoom In Zoom Out Reset image size

Repeated sequential measurement of the local PWV by the same operator allowed evaluation of the reproducibility of the measurement system. Agreement between two consecutive sets of measurement (trial I and trial II) was evaluated via Pearson's correlation coefficient (r) and Bland–Altman analysis. The statistical significance of all tests was indicated by p-value  <0.05. The mean carotid PWV from the first set of measurements did not differ significantly from the second set for each subject. As illustrated in figure 9(a), the Bland–Altman analysis shows a non-significant bias of 0.02 m s⁻¹ with an SD of the differences equal to 0.11 m s⁻¹. Agreement between the two consecutive measurements from multiple subjects is depicted in figure 9(b). The correlation between the two sets of measurements is highly significant (r  =  0.96, p  <  0.001). Study results favor the reproducibility of the developed local PWV measurement system.

Zoom In Zoom Out Reset image size

The single-source PPG device advanced with finger PPG and ECG sub-units was capable of simultaneous acquisition of the required biopotential signals. Reliable estimates of local PWV, PTT_R (beat-to-beat variation  <6.39%) and PAT_C (beat-to-beat variation  <8.01%) were obtained. The absolute value of measured PTT_R and PAT_C were in the ranges 229.16  ±  27.96 ms and 86.78  ±  12.61 ms, respectively. Mean carotid local PWV, PTT_R, and PAT_C of all recruited volunteers were consolidated along with their brachial BP for statistical analysis and to develop inter-subject mathematical models for cuffless BP measurement. Figures 10(a)–(i) show the relationship between brachial BP parameters (SBP, DBP, and MAP) versus mean local PWV, PTT_R, and PAT_C pooled over all 35 subjects. PTT_R and PAT_C were preferred over respective regional PWV values in cuffless BP models in order to avoid error due to the coarse approximation of distance between the test points (Pereira et al 2015).

Zoom In Zoom Out Reset image size

The logarithmic function provided the best monotonic fit with maximum correlation for the data points of each BP parameter versus each PTT estimate. Inter-subject mathematical models of the form BP  =  ϕ ln (υ)  +  ω (υ can be the local PWV or PTT_R or PAT_C; ϕ, ω are population-specific constants), and the absolute value of correlation coefficients (|r|) are summarized in table 1. The correlation coefficients were statistically significant (p  <  0.001) and all three transit time estimates were able to track BP parameters. However, the carotid local PWV showed better correlation (r  >  0.74) with brachial BP levels than PTT_R (|r|  =  0.62–0.67) and PAT_C (|r|  =  0.52–0.68). Inter-subject models derived from the first group of 35 subjects were used for cuffless BP estimation in a different group of 28 subjects (age  =  29.5  ±  6 years, gender: 20 males and 8 females; height  =  170  ±  9 cm; weight  =  75  ±  12 kg). Local PWV, PTT_R, PAT_C and reference brachial BP were recorded from the second group by following the same protocol discussed in section 5.3.

The cuffless BP estimation accuracy of each transit time estimate was compared with the reference brachial BP via Bland–Altman analysis as well as by computing the root-mean-square error (RMSE). Comparison study results are presented in table 1. Bland–Altman analysis showed non-significant bias (absolute value  <1.6 mmHg) for all estimated BP values. The carotid local PWV yielded good DBP estimates with a RMSE of 6.33 mmHg; it was 34% and 48% lower than the RMSE observed in cuffless DBP prediction using PTT_R, and PAT_C, respectively. The local PWV also provided 23% and 14% lower RMSE for cuffless SBP prediction than PTT_R, and PAT_C, respectively. Statistical comparison of the reference versus estimated BP parameters was made using Pearson's correlation analysis and Student's paired t-test. The difference between the reference and cuffless estimated BP in all nine cases was statistically insignificant (p  >  0.05). However, the local PWV-based BP prediction method showed better correlation with reference values (SBP: r  =  0.73, DBP: r  =  0.79, MAP: r  =  0.77; p  <  0.001). Similar to PTT_R, the PAT_C-based BP prediction method was moderately correlated with the reference BP parameters (r  =  0.59–0.66; p  <  0.05). For completeness, the conventional PAT (PAT at the fingertip (PAT_R)) acquired from the second group of volunteers was also used to predict the BP parameters. It was mapped through the PAT_R-based BP prediction models derived from the first group data. Even with the best possible calibration, PAT_R yielded RMSE  >  15 mmHg and r  <  0.55 (p  <  0.05) with the reference while estimating cuffless SBP, DBP, and MAP.

The treadmill-running exercise induced a wide range of BP and local PWV variation in individual subjects. An approximately 2.5–3.75 times increment in the carotid local PWV from the baseline value and proportional increment in BP parameters were observed after performing the exercise. Instantaneous variation in the local PWV due to arterial pressure changes during post-exercise recovery was efficiently tracked by the single-source PPG transducer. Figure 11 displays the post-exercise recovery curve (local PWV, brachial BP parameters, and heartrate as a function of time during the whole recovery period) of one subject (age  =  26 years, BMI  =  25.05 kg m⁻²). The dynamics of the curve, such as increment from the baseline value, recovery period and final steady values, were dependent on the individual's physiological factors. Based on the recovery characteristics, the post-exercise recovery curve was divided into two phases: (a) transient phase and (b) recovery phase. During the transient phase, the local PWV, BP parameters and heartrate decreased rapidly. The local PWV was scattered more in this phase due to heartrate variability, and unstable and fast breathing patterns. The local PWV and BP parameters continued to decrease at a lower rate in the recovery phase. Heartrate variability and breathing pattern were fairly stable in the second phase.

Zoom In Zoom Out Reset image size

Correlation between the intra-subject variations in the local PWV versus brachial BP parameters was obtained for each subject. The local PWV correlated equally well with SBP (r  =  0.916  ±  0.042, p  <  0.001), DBP (r  =  0.947  ±  0.033, p  <  0.001) and MAP (r  =  0.929  ±  0.025, p  <  0.001). This statistically significant tight correlation suggested developing the cuffless BP technique with intra-subject mathematical models for accurate measurement. Consistent with theory (Poon et al 2008, Mukkamala et al 2015), a logarithmic function provided the best monotonic fit for pairs of local PWV–BP data points. The developed intra-subject mathematical model was specifically of the form BP  =  Φ ln (local PWV)  +  Ω; Φ and Ω are subject-specific constants. The local PWV–BP relationship derived from the post-exercise recovery curve shown in figure 11 is depicted in figures 12(a)–(c). Similar models were developed for each subject by finding the curve that best fitted the data points obtained from their post-exercise recovery curve.

Zoom In Zoom Out Reset image size

Local PWV values collected from the recruited subjects during their routine work (second dataset) were mapped through their respective mathematical models to predict the BP level. Figures 13(a)–(c) illustrate the Bland–Altman plots of the error between the cuffless estimated and measured BP levels over all of the subjects (different symbols in the plots correspond to each of the subjects). The proposed technique was able to predict SBP, DBP, and MAP with RMSEs of 2.61 mmHg, 2.32 mmHg, and 2.34 mmHg, respectively. The corresponding RMSEs offered by mapping the local PWV through inter-subject models were 7.93 mmHg, 6.33 mmHg, and 6.74 mmHg (refer table 1). In summary, cuffless techniques based on the local PWV using inter-subject models and best-case subject-specific calibration could potentially track arterial BP parameters with a level of acceptable accuracy. By introducing more convenient methods for intra-subject model development and improving form factor (example: single-source PPG patch), the proposed single-source PPG device could be used for cuffless, ambulatory BP monitoring at clinic/home and hypertension screening test.

Zoom In Zoom Out Reset image size

The present in vivo validation study was conducted on a small population of young and healthy subjects. On the other hand, the current subjects represent an important sub-population in the sense of early detection of cardiovascular diseases. An extensive validation study with a larger population, recruited from different age groups and various pathophysiological sub-populations would confirm and extend these results. Another limitation concerns the applicability of the proposed device for self-monitoring without the assistance of a trained operator. Future efforts are required to implement fully automated intelligent algorithms in the measurement software, to guide the patient regarding probe placement and orientation to acquire dual PPG with the desired signal quality for reliable measurements. Hence, the proposed device would turn out to be portable, self-monitoring cuffless BP equipment.