Photoplethysmographic Measurement of Various Retinal Vascular Pulsation Parameters and Measurement of the Venous Phase Delay

Abstract

Purpose.: Retinal vein pulsation properties are altered by glaucoma, intracranial pressure (ICP) changes, and retinal venous occlusion, but measurements are limited to threshold measures or manual observation from video frames. We developed an objective retinal vessel pulsation measurement technique, assessed its repeatability, and used it to determine the phase relations between retinal arteries and veins.

Methods.: Twenty-three eyes of 20 glaucoma patients had video photograph recordings from their optic nerve and peripapillary retina. A modified photoplethysmographic system using video recordings taken through an ophthalmodynamometer and timed to the cardiac cycle was used. Aligned video frames of vessel segments were analyzed for blood column light absorbance, and waveform analysis was applied. Coefficient of variation (COV) was calculated from data series using recordings taken within ±1 unit ophthalmodynamometric force of each other. The time in cardiac cycles and seconds of the peak (dilation) and trough (constriction) points of the retinal arterial and vein pulse waveforms were measured.

Results.: Mean vein peak time COV was 3.4%, and arterial peak time COV was 4.4%. Lower vein peak occurred at 0.044 cardiac cycles (0.040 seconds) after the arterial peak (P = 0.0001), with upper vein peak an insignificant 0.019 cardiac cycles later. No difference in COV for any parameter was found between upper or lower hemiveins. Mean vein amplitude COV was 12.6%, and mean downslope COV was 17.7%.

Conclusions.: This technique demonstrates a small retinal venous phase lag behind arterial pulse. It is objective and applicable to any eye with clear ocular media and has moderate to high reproducibility. ( http://www.anzctr.org.au number, ACTRN12608000274370.)

Introduction

Retinal vein pulsation is a useful sign of normal intracranial pressure, and its presence is a positive prognostic sign in the assessment of glaucoma and other disorders. ^1–3 It has been considered that retinal vein pulsation timing is driven by the intraocular pressure (IOP) pulse with dilation in time with ocular diastole. ^4,5 However, recent work suggests the opposite, that vein dilation is in time with ocular and cerebrospinal fluid systole and is driven by cerebrospinal fluid pressure phase. ⁶ This suggests that retinal venous timing properties may be used to infer features concerning cerebrospinal fluid pressure. Previous work has used manual counting of video frames to assess timing characteristics. Quantitative measurements of other pulsation parameters have been limited to the threshold ophthalmodynamometric force (ODF) at the onset of pulsation, which is possible only in subjects without spontaneous venous pulsation. ^7,8 Recently, retinal vessel transverse diameter measurements have been taken, which allow transverse diameter pulsation amplitudes to be calculated. However, they can be taken only on the retinal surface and not the optic disc, where the bulk of venous pulsation is seen. ⁹  

Venous pulsation pressure (VPP) is the minimum IOP at which retinal venous pulsation is visible. This is usually determined by applying an ophthalmodynamometer to the eye, adding force to the contact lens, while an observer views the retinal vessels on the optic disc. The threshold force is converted to millimeters of mercury using a calibration factor and added to the baseline IOP to calculate the VPP. ¹⁰ Venous pulsation is generally observed on the optic disc, and it is the optic disc VPP that is associated with glaucoma severity and prognosis. ^3,7 Vein pulsation pressure is also useful in monitoring patients with central retinal vein occlusion ¹¹ and for assessing patients with altered orbital pressure from Graves disease. ² Vein pulsation pressure is linearly associated with intracranial pressure, and the presence or absence of spontaneous venous pulsation is a useful clinical sign for the assessment of elevated intracranial pressure. ¹²  

Vein pulsation pressure can be measured in only 2% to 5% of normal subjects and 46% of glaucoma subjects by ophthalmodynamometry, although in a greater proportion of subjects with raised intracranial pressure. ^7,13 It is a subjective measurement, with the “minimal detectable pulsation” threshold varying between observers. Unfortunately, the dynamic retinal vessel analyzer technique for determining dynamic vessel diameter change is not applicable to assessing or measuring vessel diameters on the optic disc. ⁹ The retinal veins on the optic disc tend to pulsate rather like the periodic flattening of a ribbon without marked transverse diameter change in many cases. Techniques restricted to diameter measurement are likely to miss some of the pulsation signal. Recent work by our group using video frame observation determined that retinal venous pulsation is broadly in phase with intraocular and intracranial pressures and dominated by the intracranial pressure pulsation. ⁶ These data suggest that intracranial pressure timing information may be inferred by the retinal vein pulse. 

Theoretical modeling of venous pulsation suggests that analysis of any pulsatile component could allow estimation of venous capacitance and resistance. ⁴ By quantifying the venous blood column pulsation, one could measure amplitude and rate (or slope) of blood column change. Both of these parameters are likely to be altered by venous resistance and may provide useful information in patients with retinal venous occlusions. Green color channel intensity in digital cameras has a peak sensitivity of 550 nanometers matching a peak in hemoglobin absorption. This creates the possibility to perform a variation of photoplethysmography ¹⁴ that has not been performed on the retina or optic nerve to our knowledge. ¹⁵ The eye is in a rather unique situation and favors this approach by being transparent and having the vessels close to the surface of the major reflecting layers, that is, retina and optic nerve. Our technique measures the transmitted light reflected back to the camera light detector. We tested and developed this system on a cohort of glaucoma patients with a particular interest in the reproducibility characteristics of several pulsation parameters and its utility in detecting pulsation phase shifts between vessel types. 

Methods

The Vein Pulsation Study trial in Glaucoma (VPSG) is a 5-year study of vein pulsation in patients being treated for and monitored for glaucoma progression. It is registered with the Australian New Zealand Clinical Trials Registry. Patients in this study routinely undergo ophthalmodynamometry and photography. Videophotography of the optic nerve and retina was performed on a subset of 20 patients in this cohort, all of whom had agreed via informed consent. The study was performed under the approval of the University of Western Australia Human Ethics Committee and adhered to the tenets of the Declaration of Helsinki. 

The patients' pupils were dilated with tropicamide 1% and had IOP as well as observation of the retinal vessels performed using a 60-diopter indirect lens at the slit lamp. The recording methodology was very similar to that reported before. ⁶ A pulse oximeter (Nellcor N65; Covidien, Mansfield, MA, USA) was applied to the right index finger. The patients had a local anesthetic (benoxinate 1%) instilled, and a Meditron (Volklingen, Germany) ophthalmodynamometer was applied. The Meditron ophthalmodynamometer uses a Goldmann three-mirror contact lens with excellent optics, and the contact lens and gel tend to dampen microsaccade and other patient eye movements, enhancing the stability of the video recording. We did attempt video recordings without contacting the eye by using a 60-diopter indirect lens; however, the quality of recordings was too poor to analyze because of patient eye movement. A video slit camera (Leica ICA, Heerbrugg, Switzerland) was used to record video footage from the patients' eyes. The slit-lamp video output was combined with audio recordings from the pulse oximeter using a sound mixer (Xenyx 802; Behringer, Willich, Germany) and the joint signal digitized to computer as a video file. The video sequences were recorded, and the applied ODF was noted for each sequence. 

Video sequences consisted of sequences of at least three cardiac cycles in length. Sequences of three cardiac cycles in length starting with the pulse oximeter signal were extracted from the video recordings along with the ODF. Pairs or triplets of recordings were included in this study where separate recordings from the same eye had been taken with the ODF within one unit of each other. There had to be at a least 30-second time interval between any pair of included recordings and at least one other force level used between any pair of recordings. 

Aligned video frames were cropped to include the optic disc and immediate peripapillary retina. The operator identified the upper hemivein, the lower hemivein, and the central part of the central retinal artery as separate vessel segments (Fig. 1). These segments were digitally dissected, extracting aligned vessel segments from each of the multiple frames. These isolated segments in multiple frames represented the time series (Fig. 2). 

Figure 1

 

Example of two video frame images with segments outlined by arrows (veins, blue; artery, red). During systole (A) the superior vein can be seen dilated, whereas during diastole (B) the superior vein is partly constricted. Corresponding measurements are in Figures 3 and 4.

Figure 2

 

Schematic diagram illustrating the sequence from raw video recording to pulse waveform production derived from aligned grayscale segmented images.

Figure 3

 

Three contiguous cardiac cycle recordings for upper hemivein (A), lower hemivein (B), and artery (C) using a standard analysis. Red, green, and blue points indicate first, second, and third cardiac cycles. Dual-frequency harmonic curve fits overlie each of these graphs. The harmonic fit over one cardiac cycle for each of the three vessel segments is shown in (D).

Figure 4

 

An empirical method using least mean squares calculates downslope (A) and upslope (B) with lines of best fit shown from portions of the pulse wave using data from three contiguous cardiac cycles. These data are separated into single cardiac cycle lengths and overlaid across a single time axis (C). A linear adjustment to baseline corrects for variation in average intensity over the whole time series (D) before the slopes are calculated. These data are from the upper hemivein data in Figure 3.

Intesity Analysis

Custom software written in R imported the separate series of segmented vessel images. ¹⁶ Each image is imported as an array and the green channel selected, creating a two-dimensional array, having pixel values between 0 (darkest) and 1.0 (lightest) green channel intensity. All images containing the vessel segment were stacked to create a three-dimensional array with time over three cardiac cycles in the third dimension. For standard analysis, the intensity was subtracted from 1.0 and multiplied by 256 to create a familiar 8-bit type of unit, with the higher value indicating greater light absorption indicative of thicker blood column. The mean of the transformed grayscale values from each image segment in the time series was analyzed. 

The optical path length through hemoglobin and its relationship to the transmission of light is described by the Beer-Lambert law, which describes the path length as proportional to the negative logarithm of the transmittance. The transmittance is the ratio between recorded light intensity to incident light intensity in photons and is related to the intensity of the light transmitted through the tissues of the eye along the optical path to the charge-coupled device (CCD) pixel element. Most camera CCD chips apply an exponential gamut function to the recorded output. We converted the raw intensity values by using the negative logarithm of the intensity value so that this value would have a stronger linear relationship to the optical path length and hence thickness of any blood vessels encountered, as well as rendering variations in camera gamut linear rather than exponential. We did calculate an approximation for measuring optical path length absorbance by hemoglobin in microns assuming an extinction coefficient of 12 L/mmol/cm at 550 nanometers for both oxy- and deoxyhemoglobin ¹⁷ and an average hemoglobin concentration of 150 g/L in our subjects, along with a molar density of 64,500 grams hemoglobin per mole. We divided this result by 2, assuming that the bulk of light passes through and is reflected back through the same vessel encountered—meaning that the optic path length is twice the actual vessel depth. This resulted in a factor of 68, which was then multiplied with the negative logarithm. This calculation was found to give values on the same order of magnitude compared to the standard analysis, making it easier for observers to compare results. This value has not been validated empirically, and the mathematical derivation is a marked simplification of the complexities of matching between hemoglobin transmittance, reflectance, and CCD green channel sensitivity profile. 

The mean of these standard or log transform values for each segment at each time point was calculated. For each segment this generated a three cardiac cycle length time series of mean values. Two forms of analysis were applied, one using a harmonic analysis and one using an empirical system. 

Harmonic regression was performed applying a two-frequency curve fit to the time series assuming cardiac cycle periodicity (Fig. 3). A linear spline term was incorporated in the regression to account for the effects of patient movement during video capture. Differentiating the fitted curves allowed maximum slope values to be calculated per cardiac cycle (cc). A conversion to real time per second (sec) was calculated knowing the frame rate at 25 frames per second (fps) An empirical algorithm was also employed, again allowing for a linear spline baseline intensity shift to account for patient movement (Figs. 4C, 4D). The mean of the maximum and minimum values was calculated to identify the peak and trough values along with their times. Linear regression was applied to the up and down phases of these data sets to calculate downslope and upslope (Figs. 4A, 4B). 

Statistical Analysis

The raw data are expressed as mean ± standard deviation. For each calculated parameter, the mean and standard deviation was calculated for each pair or triplet video sequence with approximately equal ODF (±1 unit). A P value less than 0.05 was considered significant. Each dataset was analyzed in four ways, with the grayscale analysis by logarithmic or standard technique and the wave form by empirical or harmonic analysis, generating four measurements for each parameter type. The coefficient of variation (COV) was used as our measure of repeatability for each parameter, which we termed precision. It was calculated by standard deviation divided by mean multiplied by 100, as a percent. Box plots were produced to allow for comparison of the COV distributions between each parameter for both the artery and vein segments. Multiple measurements were taken from individuals and from the right and left eyes from the same individual. Intraclass correlation coefficients were calculated for within individual and within eye in individual. These were found to be elevated in amplitude (>0.34), downslope (>0.24), and upslope (>0.46) for both eye and individual. So for all comparison testing, linear mixed models were used, incorporating random factors to account for intraeye and intraindividual correlation. To compare different parameter COVs, these linear mixed models were produced with the COV as the dependent variable and type of segment or analysis as the explanatory variable, with patient identity and eye as nested random factors to account for multiple observations from patients and eyes. Coefficient of variation was compared to the mean values, using a similar linear mixed model, to examine whether lower COV was associated with higher mean values, that is, whether standard deviation remained relatively constant. All model residuals were then tested for normal distribution using quantile–quantile distribution; then a Box-Cox analysis and transform were applied when non-normal. ¹⁸ With non-normal distributions, optimum λ was selected and the data transformed as follows: transformed COV = (COV ^λ − 1)/λ. Then the transformed data were retested for normality. One-way analysis of variance was used to compare between factors (vessel segment or analysis type) and linear modeling for the comparison between continuous variables (COV and mean values). For each segment, six comparisons were performed, two for each of the amplitudes, downslopes, and upslopes, so using a Bonferroni correction a significance value of (0.05/6) 0.0083 was used. 

The timing of the peak and trough phases was calculated using the harmonic peak and trough calculations. The difference between arterial and upper vein times and arterial and lower vein times was calculated for each image series. These differences were analyzed using a linear mixed model with time difference as a dependent variable and artery to (upper or lower) vein difference as the explanatory variable, with patient identity and eye as nested random factors to account for multiple observations from patients. 

Results

There were 268 vessel segments analyzed from 23 eyes from the 20 patients (15 females and 5 males) with mean age 66.1 ± 9.5 years. The mean IOP of the 23 eyes tested was 14.7 mm Hg (SD 2.5) with an average visual field mean deviation of −4.4 dB (SD 5.0). Within those eyes, 34 hemiveins were spontaneously pulsating and 12 were not. In those that were not, the mean additional ODF to induce venous pulsation was 17.6 g (SD 10.6). This force can be converted to additional IOP applied through the use of a conversion constant (0.89 mm Hg/g) so that ophthalmodynamometric pressure (ODP) is equivalent to ODF × 0.89. ¹⁹ Here, the mean ODP in the eyes that did not spontaneously pulsate was 15.7 mm Hg (SD 9.4). Final datasets consisted of 89 similar pairs of data (in terms of ODF) and 30 similar triplets of data from which standard deviations, means, and COVs were calculated. Forty datasets were arterial segments; 40 were upper vein and 39 were lower vein segments. Tests of normality analyzing amplitude and slope COVs showed significant skew to the right with Box-Cox analysis prompting the use of λ = 0.3. One-way analysis of variance comparing the transformed COVs for all parameters between the upper vein and lower vein showed no significant difference, with a minimum P value of 0.38. So, COVs from the upper and lower veins were combined (the combination was referred to as veins). The distributions of the COVs for logarithmic values of harmonic amplitude in the artery and vein pulsation are shown in Figure 5 as a frequency histogram. The average of all the means and standard deviations for the pairs and triplet datasets were calculated and are presented in the Table. The mean and median COVs are also presented in the Table. 

Figure 5

 

Frequency histogram of coefficients of variation for amplitude calculated by harmonic analysis using the log transformed data of arteries (coarse dark backward slash) and veins (fine light forward slash).

Table.

 

Average of the Means and Standard Deviations Measured From the Sets of Data From Approximately Equal ODF Settings With the Calculated Coefficients of Variation (These Data Were Analyzed Using the Harmonic Regression Technique)

Vessel	Pixel Analysis	Parameter	Average Mean	Average SD	COV Mean, %	COV Median, %
Vein	Logarithmic	Amplitude	7.5	0.96	12.6	10.3
	Standard	Amplitude	9.9	1.16	12.7	11.2
	Logarithmic	Downslope cc	−22.0	4.00	17.7	14.4
	Standard	Downslope cc	−28.6	4.60	16.0	12.4
	Logarithmic	Upslope cc	32.7	4.16	12.8	10.7
Artery	Logarithmic	Amplitude	4.8	0.62	12.7	11.7
	Standard	Amplitude	5.3	0.70	12.1	9.3
	Logarithmic	Downslope cc	−14.3	2.60	18.4	17.2
	Standard	Downslope cc	−16.2	3.02	17.9	16.5
	Logarithmic	Upslope cc	21.0	3.14	15.2	13.6

Figures 6 and 7 demonstrate the box plots for the COVs for both the venous and arterial parameters. One can compare the overall distributions and medians between the vessel types as well as between the standard and logarithmic analyses. No differences were found between the COVs for the logarithmic and standard analysis (minimum P value of 0.61) for either veins or arteries. This suggested that there was no significant difference in terms of repeatability between logarithmic and standard analysis. 

Figure 6

 

Box plots of the coefficients of variation for the key parameters measured in the vein segments. Std, standard pixel analysis; log, logarithmic pixel analysis; emp, empirical waveform analysis; harm, harmonic waveform analysis; cc, time in cardiac cycles; sec, time in seconds.

Figure 7

 

Box plots of the coefficients of variation for the key parameters measured in the artery segments. Std, standard pixel analysis; log, logarithmic pixel analysis; emp, empirical waveform analysis; harm, harmonic waveform analysis; cc, time in cardiac cycles; sec, time in seconds.

The mean values did not vary significantly between the harmonic and empirical results analyzing amplitudes for any vessel (minimum P = 0.21). However, the absolute magnitudes of down and up slopes were greater in the harmonic group compared to the empirical group for all vessels (all P < 0.0001). Between the harmonic and empirical analysis for any parameter in the venous segments, the COV did not vary significantly (minimum P = 0.33). Comparing between harmonic and empirical analysis for any parameter in the arterial segments revealed no significant difference in COV for any parameter (minimum P = 0.039 comparing the logarithmic downslope per cardiac cycle). 

No significant relationship could be found between COV and mean results for any logarithmic parameter in either arteries or veins (minimum P = 0.32). Likewise, the COV was not associated with mean results for any standard parameter in either arteries or veins (minimum P = 0.063 for vein upslope per second). 

The vessel pulsation timing was analyzed in terms of the peaks (peak light absorption corresponding to dilation) and troughs (minimum light absorption corresponding to constriction). Figure 8 shows the frequency histogram of COVs for both the vein and artery peak times. One can see a marked skew to the left with a heavy concentration of very low COVs, suggesting that generally there was high precision for determining the time of peak. 

Figure 8

 

Frequency histogram of coefficients of variation for time at vessel peak (dilation) calculated by harmonic analysis using the log transformed data of arteries (coarse dark backward slash) and veins (fine light forward slash).

The mean vein peak time in terms of cardiac cycle mean was 0.88 cardiac cycles ± 0.03 with a mean COV of 3.4% and median of 2.0%. The arterial peak time mean was 0.81 cardiac cycles ± 0.03 with mean COV of 4.4% and median of 2.9%. We analyzed the upper vein to arterial peak and trough time difference and lower vein to arterial peak and trough time difference in order to determine any potential phase shift between the retinal arteries and veins. Lower vein peak occurred at 0.044 cardiac cycles after the arterial peak (P = 0.0001), with upper vein peak occurring an insignificant 0.019 cardiac cycles after the lower vein peak (P = 0.15). In terms of absolute times, the lower vein dilation occurred at mean 0.040 seconds after the arterial peak (P = 0.0005) with the upper vein occurring an insignificant 0.02 seconds after the lower vein (P = 0.15). The venous pulse trough occurred at 0.064 cardiac cycles after the arterial trough (P = 0.0000). The upper vein trough occurred an insignificant 0.007 cardiac cycles after the lower vein (P = 0.25). In terms of absolute time, the venous constriction occurred 0.062 seconds after the arterial (P = 0.0000), with the upper vein occurring an insignificant 0.005 seconds after the lower vein (P = 0.34). 

Discussion

An advantage of this photoplethysmographic technique is that it is applicable to all eyes with clear ocular media because it is not a threshold measurement. It involves relatively straightforward mathematical treatment of image densitometric values and benefits from the fact that hemoglobin is the major pigment and variations in its absorbance are indicative of the state of vessel collapse or dilation. The units of amplitude presented here should be viewed as arbitrary units. In particular, the logarithmic units, while they may be related to vessel thickness, should not necessarily be viewed as such. The mean COV for logarithmic analysis of vein amplitude (13%) and mean downslope COV (16%) are comparable to ophthalmodynamometer reproducibility figures. ^10,20 The very low (4%) COV found for peak time indicates a high degree of accuracy for detecting the timing of pulsation phase. It should be noted that the COV distribution was very similar to that reported for ophthalmodynamometry, ²⁰ and did reach 50% for vein amplitude (Fig. 5) in one individual and 28% for peak time (Fig. 8) in another individual. The technique gained best results in subjects with clear ocular media and at present appears best suited to those subjects. 

Some variability is to be expected given the nature of ophthalmodynamometry, especially since all videos taken so far have used an ophthalmodynamometer partly to dampen extraocular movements. The mean COVs for Meditron ophthalmodynamometer measurements are from 8.5% ¹⁰ to 16.3%, ²⁰ and it is possible that this inherent variation is a significant component of the variability seen in our technique by adding an element of force and hence IOP variation to the eye. 

There was no difference in the COV between the logarithmic analysis and standard analysis. The logarithmic values may be more representative of vessel thickness and may have more applicability, although this is yet to be validated. There was also no difference in COV found between the harmonic and empirical analysis in venous segments. 

An example of the applicability is the identification of a small but significant delay between venous and arterial peak and troughs. This indicates a small phase shift of approximately 0.05 cardiac cycles, with the venous phase following the arterial phase corresponding to a very slight phase lag of approximately 0.04 seconds. This is in broad agreement with our earlier work finding that vein pulsation is essentially in phase with IOP and cerebrospinal fluid pressure. ⁶ Most observers have assumed that arterial pressure and IOP are roughly in phase, with a slight lag in IOP and cerebrospinal fluid pressure ²¹ after arterial pressure. The phase of retinal venous pulsation appears to be driven by cerebrospinal fluid phase, ⁶ which suggests that further studies of phase changes in normality or in disease states may give useful information concerning cerebrospinal fluid pressure and phase. The high reproducibility of this technique makes it possible to detect small phase shifts, which may also be useful in retinal venous occlusion. Additionally, the ability to measure venous pulsation amplitude with reasonable precision may allow useful estimation of flow resistance in patients with or at risk for retinal venous occlusion. 

One limitation of this study is that it was done on a group of glaucoma patients and may not be representative of normals. However, we have found similar phase relationships between IOP and vein pulsation in glaucoma patients and subjects with elevated intracranial pressure, suggesting that a variation in disease state may not alter phase relationships. ^6,22 The exact relationship between the intensity measurement and column thickness is not known; however, this will not alter the timing relationships. It is not known whether the detection of such small phase shifts will be clinically useful. Another limitation of this technique is that a contact lens is required to gain images of sufficient quality needed for analysis. 

A major advantage of this technique is its near universal applicability and the fact that it measures pulsatility features from the optic disc. The amplitude or slope measurements may give a useful measure of pulsatility. Further analysis of these waveforms may help determine vascular resistance, capacitance, or other useful parameters of vessel function. They may be applicable to the diagnosis and monitoring of certain vascular and nonvascular diseases such as retinal vein occlusions, glaucoma, disorders involving raised intracranial pressure, and orbital disease. 

Acknowledgments

Supported by National Health and Medical Research Council Project Grant 102367 and Development Grant 107310. 

Disclosure: W.H. Morgan, P; M.L. Hazelton, None; B.D. Betz-Stablein, None; D.-Y. Yu, P; C.R.P. Lind, None; V. Ravichandran, None; P.H. House, None