2.1.

Clinical Design

A sample population of eight ESRD patients undergoing regular hemodialysis treatment ( $4\text{-}\mathrm{h}$ sessions, three times a week) was recruited on a volunteer basis for the pilot study. The use of human subjects, extraction of blood samples, informed patient consent, and confidentiality requirements were all approved by the Research Ethics Board of the Ottawa Hospital. Candidate volunteers were identified by the Kidney Research Centre staff from the population of patients treated at the Hemodialysis Unit of the Ottawa Hospital General Campus. Volunteers represented a broad cross section of ESRD patients in terms of age and gender (6 male, 2 female; mean age 61.5 years; age range 39 to 75 years); time since initiation of dialysis (18 to 284 months); and the presence of other systemic conditions (3 hypertensives, 3 Type-II diabetics). Blood was extracted from the vascular access point and occurred immediately before and after a single hemodialysis treatment $(<1\min )$ . The day of blood extraction coincided with the monthly laboratory blood testing day for each volunteer, thereby allowing subsequent correlation of spectral data to clinical laboratory results.

2.2.

Blood Sample Preparation

Samples of whole blood for the study were obtained at the same time the standard clinical blood samples were obtained. Blood was drawn into standard $3\text{-}\mathrm{ml}$ purple-top collection tubes containing $5.4\mathrm{mg}$ ${\mathrm{K}}_2$ -ethylenediaminetetracetic acid as an anticoagulant. For spectroscopy measurements, blood samples were brought to the Visual Optics Laboratory of the University of Ottawa Eye Institute, colocated with the Ottawa Hospital General Campus. The collection tubes were manually agitated to provide a homogeneous suspension and $1\mathrm{ml}$ from each tube was transferred to an optical cell and sealed. The delay between sample collection and the start of optical measurements averaged $1\mathrm{h}$ , with a maximum delay of $2\mathrm{h}$ . To test the influence of the delay, optical spectra from a single blood sample were taken hourly over a $4\text{-}\mathrm{h}$ period at room temperature, with no significant change observed in the spectra (data not shown). Clinical blood samples were sent to the clinical laboratory at the Ottawa Hospital for standard clinical evaluation. Serum urea and potassium levels were quantified using an automated Beckman Coulter LX20 analyzer. Manufacturer-supplied reagents were used and an indirect ion selective electrode method was used to quantify potassium, while a coupled enzymatic rate method was used for urea. The intra-assay variability of these techniques is nominally accepted to be approximately 2% (coefficient of variation).

2.3.

Measurement of Transmission and Diffuse Reflection Spectra

Optical cells with a $2\text{-}\mathrm{mm}$ optical path length in air were used, made from optical glass with $>80\%$ transmission over the wavelength range 365 to $2500\mathrm{nm}$ (Varsal Inc.). For measurements, the optical cells were placed in a custom sample holder ensuring both repeatable cell placement and minimal optical-mechanical interference to avoid stray light and spurious reflections. A focused spot ( $4\text{-}\mathrm{mm}$ diameter) from a current-stabilized $20\text{-}\mathrm{W}$ tungsten-halogen light source (model ASB-W-020, Spectral Products, Inc.) was used to illuminate the optical cell. Light transmitted through the cell was focused into a collection optical fiber ( $400\text{-}\mathrm{\mu }\mathrm{m}$ core diameter low-OH fiber, Ocean Optics Inc.) connected to one of two spectrometers. Light backscattered in a cone over a 10- to $30\text{-}\mathrm{deg}$ angle relative to the incident beam was captured by a large-aperture achromatic lens and focused to a second collection optical fiber ( $600\text{-}\mathrm{\mu }\mathrm{m}$ core diameter low-OH fiber, Ocean Optics Inc.) connected to the other spectrometer. The off-axis geometry used for backscatter collection minimized the interference of both specular reflections and edge effects originating from the optical cell geometry.

Optical spectra were acquired over the 400- to $1700\text{-}\mathrm{nm}$ region using two spectrometers spanning wavelength ranges of 400- to $1000\text{-}\mathrm{nm}$ (model SD2000, Ocean Optics Inc., 2048-element silicon photodiode array; spectral resolution $0.33\mathrm{nm}$ ) and 900- to $1700\text{-}\mathrm{nm}$ (model $\mathrm{InGaAs}512$ , StellarNet Inc.; 512-element $\mathrm{InGaAs}$ photodiode array; spectral resolution $2.25\mathrm{nm}$ ). Spectra were acquired through computer control with acquisition times of $8\mathrm{ms}$ and $800\mathrm{ms}$ for transmission and diffuse reflection, respectively, for the near infrared spectrometer, and $10\mathrm{ms}$ for both modes using the visible spectrometer.

Blood samples were maintained at room temperature and sample heating was minimized by using a mechanical shutter to limit sample exposure by keeping it open only during the spectral acquisition period. Three data sets were obtained for each sample, where for each set the cell was removed, agitated, and replaced. The three spectra were subsequently area normalized and then averaged. Prior to averaging, the maximum coefficient of variation among any set of three normalized spectra was 2 and 9% for the visible and near infrared spectrometers, respectively. All subsequent analyses were performed using normalized mean spectra, smoothed with a 10-point moving average filter. Influences of the spectral properties of the light source, the optical cell, and optical elements in the light delivery and detection paths were removed by dividing the transmission and diffuse reflection spectra by a reference measurement taken with an empty optical cell (transmission path) and a broadband mirror placed behind an empty optical cell (diffuse reflection path).

About $1\mathrm{W}$ of focused optical power was delivered to the blood sample. Typically about 15% of the incident light was transmitted through a sample while about 5% was diffusely reflected in the direction of the detection cone. Although the transmitted and diffusely reflected light levels were high, signal-to-noise ratio was reduced due to manual attenuation of the delivered and/or detected light streams, which was necessary to accommodate both a limited photodetector dynamic range and a requirement for the simultaneous measurement of transmitted and diffusely reflected paths with differing light levels. Wavelength regions where optical signal-to-noise levels failed to exceed an imposed $3\text{-}\mathrm{dB}$ minimum threshold were identified and excluded from subsequent analyses. Of particular note, strong water absorption in the 1400 to $1550\text{-}\mathrm{nm}$ band resulted in a reduced signal-to-noise ratio.

3.1.

Whole Blood Spectra

32 mean spectra were obtained; these corresponded to pre- and postdialysis blood for eight patients in both transmission and diffuse reflection modes. The spectra are given in absorbance units in Figs. 1a and 1b , defined as the base 10 logarithm of detected intensity relative to the reference intensity level. The spectra differed depending upon the light interaction mode, which is consistent with results reported by others.¹⁵ High absorption at wavelengths below $600\mathrm{nm}$ in Fig. 1a was a result of the relatively long $2\text{-}\mathrm{mm}$ path length through blood, while in Fig. 1b, diffusely reflected light originating from superficial blood layers underwent less attenuation thereby revealing the characteristic double-peaked absorption of oxyhemoglobin (peaks at 542 and $575\mathrm{nm}$ ). In general, excellent agreement is observed between the spectrum in Fig. 1b and the published absorption spectrum of pure oxyhemoglobin.¹⁶ Hemoglobin (principally its oxygenated and deoxygenated forms) dominates whole blood absorption of wavelengths shorter than $1000\mathrm{nm}$ , whereas water generally dominates absorption above $1000\mathrm{nm}$ . The well-known absorption peak of water centered at $1440\mathrm{nm}$ and a minor peak around $1200\mathrm{nm}$ are also visible in the measured spectra in Fig. 1.¹⁷

Fig. 1

Pre-and postdialysis whole-blood spectra in absorbance units for (a) transmission and (b) diffuse reflection modes. Blood spectra from individual patients are in gray, while average pre- and postdialysis values are indicated by dashed and solid black lines, respectively. Shaded regions of the abscissa indicate wavelength ranges with signal-to-noise ratio $<3\mathrm{dB}$ . The discontinuity around $1000\mathrm{nm}$ is due to a detected intensity mismatch between the two spectrometers, indicated by a vertical dashed line.

3.2.

Principal Component Analysis

To assess the significance of spectral changes observed as a result of hemodialysis, quantitative analysis was performed using the principal component analysis (PCA) method.¹⁸ Briefly, in the PCA method, a group of spectra are mathematically decomposed into a small set of uncorrelated, orthonormal variables (the principal components) that account for the major sources of variation across the group. For a given spectrum, it is then possible to derive a set of principal component “scores” or “weights” representing the contribution of each principal component to the linear decomposition of that spectrum in terms of the principal components. By limiting the analysis to only the most significant sources of variation among the spectra, an entire optical spectrum can be represented by one or a few variables.

PCA was performed three times (for transmission, diffuse reflection, and concatenated transmission $+$ diffuse reflection), in each case using the 16 spectra (pre- and posttreatment for 8 patients) shown in Fig. 1. In each analysis, the first four principal components accounted for nearly all the variation $(>97\%)$ among the spectra, with the first principal component representing the dominant source of variation (Table 1 ). Scores for the first principal component for each patient sample are given along with an aggregate score for the first four principal components in Table 2 . The aggregate score was computed as the sum of the squared scores of the first four principal components. Mean scores for each group (pre- and postdialysis) are shown along with the corresponding results of a paired $t$ -test (two-tailed significance $\alpha =0.05$ ) for the null hypothesis (no spectral difference due to treatment). For transmission, diffuse reflection, and combined modes, the null hypothesis could be rejected using scores from both the first and the first four principal components, indicating that hemodialysis treatment induced significant whole-blood spectral changes in the patient group. Moreover, significance at the high level of explained variance indicates that the observed spectral changes are almost completely due to the effect of the hemodialysis treatment alone.

Table 1

Percent of the total variation in optical spectra among the 16 blood samples explained by a given principal component, for the different light interaction modes.

Table 2

Principal component scores used to test the null hypothesis of no difference between pre- and postdialysis blood based on full-spectrum (500 to 1700nm ) analysis.

3.3.

Difference Spectra

Optical difference spectra (posttreatment $-$ pretreatment) were calculated from the data shown in Fig. 1 and are given in Fig. 2 . In the difference spectra, coinciding with local extrema are observed as well as isobestic wavelengths of near-zero change for all patients, which differ slightly in location depending on the light interaction mode. The magnitude of spectral change also differed among patients and was mode-dependent. For example, for first three local extrema in transmission [Fig. 2a], patient 6 exhibited the largest change while patient 4 exhibited the largest change for the following two extrema. In diffuse reflection [Fig. 2b], however, patient 2 shows the largest change at the first minimum while patient 3 shows the largest change at the following peak and patient 4 shows the largest change at the next minimum.

Fig. 2

Whole blood difference spectra (postdialysis $-$ predialysis) for eight patients in (a) transmission and (b) diffuse reflection modes. A horizontal line at zero change is included for reference.

Periodic fluctuations observed in the diffuse reflection change at longer wavelengths are due to multiple reflections caused by the blood, optical cell, and air interfaces. Moreover, only a portion of the diffusely reflected light ( $20\text{-}\mathrm{deg}$ solid angle) was captured in our setup resulting in a low absolute intensity level and greater relative noise contribution. It has been reported that increasing the solid angle of collection of diffusely reflected light using an integrating sphere significantly improves the quality of the whole-blood spectra.¹⁹

3.4.

Correlation with Clinical Measures

We sought to investigate whether the observed whole-blood spectral changes (expressed as difference spectra) were consistent with known blood chemistry changes due to hemodialysis therapy. For each patient, the clinical record of the hemodialysis session was used to compile key measures of hemodialysis performance to allow a direct comparison of optical measurements and clinical outcomes. The analysis presented here focuses on a few key clinical measures of hemodialysis treatment: the urea reduction ratio ( $\mathrm{URR}=1-$ post-preblood urea concentration); a derived measure, the potassium reduction ratio ( $\mathrm{KRR}=1-$ post-preblood potassium ion concentration); a derived measure relating to fluid removal by filtration, the weight retention ratio ( $\mathrm{WRR}=\mathrm{post}\text{-}\mathrm{prebody}$ weight); and $Kt∕V$ (dialysis dose, where $K$ is urea clearance rate of the dialyzer in liters per minute, $t$ is treatment time in minutes, and $V$ is the urea distribution volume for the patient in liters).

The clinical values for the chosen measures along with the range of values among the patient group are shown in Table 3 . The value of Pearson’s correlation coefficient $r$ among the clinical parameters has also been given in Table 3. A nearly perfect positive correlation was found between URR and $Kt∕V$ , which was expected given that $Kt∕V$ values are derived from the URR of a patient.²⁰ URR thus suffices to describe the behavior of $Kt∕V$ in the following analysis. Furthermore, while absolute potassium reduction is used as an accepted clinical measure, KRR was derived for the purposes of this study to provide a relative, unitless parameter for consistency in the analysis. Absolute potassium reduction and the KRR were highly correlated $(r>0.97)$ indicating nearly identical behavior.

Table 3

Hemodialysis parameter values for the patient group obtained from clinical blood laboratory results and patient charts. Minimum and maximum parameter values within the group (predialysis) are also given. Correlations among the parameter values are also shown. Parameter definitions are given in the text.

Additionally in Table 3, correlations were observed between WRR and KRR as well as URR and KRR. To remove the effect of intervening parameters, the partial correlation coefficient $r_{xy,z}$ has been used to represent the correlation of $x$ (the set of optical difference values at a given wavelength data point from Fig. 2) with $y$ (the set of clinical parameter values), while removing the influence of $z$ (a correlated clinical parameter). The result is a partial correlation spectrum for each clinical parameter with the influence of the most confounding clinical parameter factored out. The partial correlation spectra for the WRR, URR, and KRR are given in Fig. 3 . Correlation spectra from the optical interaction mode (transmission or diffuse reflection) with the greatest degree of variation for each parameter are shown. While a detailed interpretation of wavelength regions of high correlation is precluded by the small sample size, general observations concerning the overall nature of the variations and their likely origins can be made.¹⁸ In particular, the presence of large correlation variations with wavelength indicates the influence of molecular absorption bands.¹⁸ Second- and third-order overtones of the $\mathrm{N}–\mathrm{H}$ symmetric and asymmetric stretching vibrations in the urea molecule are broad bands centered at 960 to $1000\mathrm{nm}$ and 720 to $750\mathrm{nm}$ , respectively.⁶ First-order overtones of these stretches occur in the 1450- to $1500\text{-}\mathrm{nm}$ region, but in blood samples these would be buried under the strong water absorption band in this region.⁶ Additionally, potassium ions in whole blood have been shown to exhibit a broad correlation with absorption in the 500- to 1000-nm wavelength region.²¹ Observed regions of high correlation may correspond to analyte levels directly or may be due to more subtle, indirect effects related to their removal. For instance, potassium is known to play a role in the intra- and extracellular fluid balance, potentially influencing blood cell volume and thereby the diffuse reflection spectrum. In addition, the WRR (or weight change, a measure of fluid removal during hemodialysis) reflects not only a hemoconcentrating effect in the blood, but relates indirectly to the removal of a number of water-soluble solutes during treatment—effects that may contribute to the observed spectral changes.

Fig. 3

Partial correlation spectra for (a) WRR with the effect of KRR removed, (b) URR with the effect of KRR removed, and (c) KRR with the effect of WRR removed. The spectra represent partial correlation with the measured transmission or diffuse reflection difference spectra from Fig. 2.