- Email: [email protected]
Preliminary investigation of near-infrared spectroscopic measurements of urea, creatinine, glucose, protein, and ketone in urine
Clinical Biochemistry 34 (2001) 239 –246
Preliminary investigation of near-infrared spectroscopic measurements of urea, creatinine, glucose, protein, and ketone in urine J. Larry Pezzaniti, Tzyy-Wen Jeng, Larry McDowell, Gary M. Oosta* Abbott Laboratories, Diagnostic Division, Abbott Park, IL 60064, USA Received 14 November 2000; received in revised 14 February 2001; accepted 16 February 2001
Abstract Objective: We investigated the use of near-infrared spectroscopy as an analytical tool to quantify concentrations of urea, creatinine, glucose, ketone, and protein in urine. Design and Methods: FT-IR spectroscopy in conjunction with a polynomial based spectral smoothing method was applied to urine specimens. A partial factorial experimental design was employed to collect spectra using normal and spiked urine samples. Results: Our results show that the spectral signatures of urea, creatinine, glucose, ketone, and protein in the 1350 to 1800 nm and 2050 to 2375 nm range are sufficiently strong and unique for accurate measurements. Conclusions: The accuracy of near infrared for quantifying concentrations of urea and creatinine is only slightly less than our selected reference methods. Glucose, ketone and protein are sufficiently accurate to be useful as a screening tool for wellness. The method successfully accounts for biologic matrix variation. The advantages of near-infrared analysis are (1) no reagents, (2) ease of sample preparation, (3) speed, and (4) the ability to quantify multiple analytes with one spectra. © 2001 The Canadian Society of Clinical Chemists. All rights reserved. Keywords: Urinalysis; Urine; Spectroscopy; Infrared; Near-infrared; Visible; Fourier transform
1. Introduction For clinical laboratory measurements, quantitative methods must be suitably accurate and precise over the expected range of values required. In addition, it is often desirable that the method be inexpensive, reliable, rapid and easily automated. Near-infrared spectroscopy has the potential to satisfy these criteria. It needs no reagents, little or no sample preparation, it is rapid and nondestructive, and is suitable for complex matrices. Near-infrared spectroscopy has been applied to measuring urine composition [1,2], serum composition [3,4], fecal composition [5,6], glucose in whole blood [7] and complex matrices [8]. The paper explores the feasibility of the use of near-infrared analysis for measuring five compounds of interest for urinalysis screening testing. Urine contains a wide variety of substances. Urinalysis involves measuring critical components in a sample of urine to identify previously undetected diseases or medical con* Corresponding author. Tel.: ⫹1-847-937-2553; fax: ⫹1-847-9386929. E-mail address: [email protected] (G.M. Oosta).
ditions, or urinalysis may be used to determine if regulated substances (e.g., drugs) are being abused. All of these analytes represent breakdown products of metabolism from various organ systems. The pattern of excretion is indicative of various disease states. The history and utility of urinalysis have been reviewed [9,10]. In current urinalysis systems, such as those provided by Bayer and Boehringer Mannheim, the analytes measured include glucose, bilirubin, blood (or hemoglobin), protein, urobilinogen, nitrites, leukocytes, specific gravity, and pH. In urine the major ketone components are 3-hydroxybutyrate (80%), acetoacetic acid (17%) and acetone (3%), but only the acetoacetic acid is determined by the current test systems. Refractive index may be substituted for specific gravity [11,12]. In some cases, measurement of creatinine is suggested, but is not provided by Bayer’s or Boehringer Mannheim’s urinalysis systems [13,14]. The majority of urinalysis testing is accomplished by means of dip and read strip technology supplied by Bayer and Boehringer Mannheim. Strip technology is well understood, and suffers from a number of limitations. Readings must be properly timed to obtain accurate results. Urine
0009-9120/01/$ – see front matter © 2001 The Canadian Society of Clinical Chemists. All rights reserved. PII: S 0 0 0 9 - 9 1 2 0 ( 0 1 ) 0 0 1 9 8 - 9
240
J.L. Pezzaniti et al. / Clinical Biochemistry 34 (2001) 239 –246
samples must be well mixed and at room temperature. Strips are sensitive to light and humidity, and must be stored and handled properly. Quantitative results are difficult to obtain. Interfering substances can cause incorrect readings. This work demonstrates the utility of near-infrared analysis for measurement of glucose, ketone and protein (currently done with test strips), and urea and creatinine. The objective was to determine the accuracy of near-IR spectroscopy for determining concentrations of these analytes in urine.
2. Experimental protocol and instrumentation Two experimental protocols, referred to as protocols 1 and 2, were used to prepare the urine samples and collect the NIR data. For measurements involving urea, creatinine, glucose, experimental protocol 1 was employed. For protein and ketone, both experimental protocols 1 and 2 were used. Protocol 1 was designed to produce samples of varying analyte concentration and to provide a variable matrix from which an accurate partial least squares (PLS) model could be derived and to provide an independent set of samples that could be used to check the accuracy of the PLS model. Protocol 1 consisted of the following steps. A. First morning urine specimens were collected from sixty different healthy donors. B. The samples were filtered with a 0.45 micron pore size cellulose acetate filter (Nalge Nunc International, Inc.). C. Each specimen was divided into two portions to make two identical sample sets, samples 1 to 60 and samples 61 to 120. D. Samples 1 to 60 were used without change. E. Additional analytes were added to samples 61 to 120 according to a randomization protocol. In some cases, two or more analytes were added to the same sample. The randomization process provided samples that gave the high and low values needed for a calibration sample set and the high and low test sample concentrations that were needed for a validation sample set. In protocol 1, the final analyte concentrations were within the range of concentrations that might be expected in a clinical laboratory. Even numbered samples (1–120) were used as a validation set to check the accuracy of the PLS model that was developed. F. Odd numbered samples (1–120) were used as the calibration set for the development of the PLS model. Figure 1 shows the distributions for the concentrations of the analytes in samples 1 to 120 used for protocol. Protocol 2, while similar to protocol 1, incorporated the following changes. 1. Thirty first morning urine samples were used without prefiltration 2. When analytes were added as described in step E, the resulting concentrations were substantially greater than might be expected in a clinical laboratory
Fig. 1. Histogram showing the distribution of concentrations of urea, creatinine, glucose, protein and ketone in calibration and model validation samples, for protocol 1.
Figure 2 shows the distributions for the concentrations of the analytes in samples used for protocol 2. A sample handling system was used to deliver the samples to the cuvette that was held at a constant temperature of approximately 25.0 C. An automated Gilson 223 Sample Changer (Gilson Inc.) was used in conjunction with a constant temperature sample delivery system. The sample delivery system consisted of a 100 cm of 0.076 cm inside diameter aluminum tube submerged in a water bath maintained at 25.0 C ⫾ 0.01. The sample was be pumped through the aluminum tube with a peristaltic pump such that the sample would arrive at the cuvette at the desired temperature. The cuvette was held in thermal contact with a water-jacketed cuvette holder also maintained at 25.0 C ⫾ 0.01. The standard deviation of the measured sample to sample temperature was less than 0.02 C, as measured with a k-type thermocouple in contact with the urine. Carry-over
Fig. 2. Histrogram showing the distribution of protein and ketone in sample sets spiked with analyte beyond the normal range for healthy urines, for protocol 2.
J.L. Pezzaniti et al. / Clinical Biochemistry 34 (2001) 239 –246
of one sample to the next was reduced to less than 0.05% by a wash protocol between samples. Near-IR spectra were collected with an Equinox 55/S FTIR spectrometer (Bruker Inc.). The settings of the spectrometer where optimized for NIR measurements by using its 100W tungsten-halogen lamp, its Indium Galium Arsenide thermoelectrically cooled detector, and its quartz beam splitter. The spectral range of the measurements was 1100 to 2500 nm. The cuvette was a quartz flow through cell with a pathlength of 1 mm. To achieve a satisfactory signalto-noise ratio, 32 spectra were averaged at a spectral resolution of 2 cm⫺1
3. Sample preparation and reference methods Fresh urine samples from normal, healthy volunteers were collected for initial calibration and model validation testing. Urea (Aldrich Chemical Co., Milwaukee, WI) powder was weighed out to the nearest 0.1 mg, and dissolved into 10 ml of sample so that the sample set’s concentrations were of sufficient dynamic range for calibration. Urinary Urea Nitrogen concentration of both native urine and modified urine was determined by the Vision Instrument (Abbott) after predilution with distilled water. Appropriate dilution factor (usually between 51–26 fold) was used so that the diluted sample would contain urea within the normal range of the Vision Urea Nitrogen Assay which had been developed and optimized for serum. Creatinine hydrochloride (Sigma, St. Louis, MO) was weighed out to the nearest 0.1 mg, and dissolved into 10 ml of sample so that the sample set’s concentrations were of sufficient dynamic range for calibration. Urinary creatinine concentration of both native urine and modified urine was determined by the Vision Instrument (Abbott) after predilution with distilled water. Appropriate dilution factor (usually between 33–26 fold) was used so that the diluted sample would contain creatinine within the functional range of the Vision Creatinine Assay which had been developed and optimized for serum. 3-Hydroxybutyric Acid (Aldrich Chemical Co., Milwaukee, WI) powder was weighed out to the nearest 0.1 mg, and dissolved into 10 ml of sample so that the sample set’s concentrations were of sufficient dynamic range for calibration. An enzymatic method for assaying -hydroxybutyrate from Sigma Diagnostics (St. Louis, MO) was used to measure the concentration of ketone in both the native and spiked urine samples. BSA Stock solution was prepared by dissolving 40 g of bovine serum albumin (BSA) powder (Fraction V, Sigma Chemical Co., St. Louis, MO) per liter. It was stored at 4°C. The Micro-protein Assay (Sigma Diagnostics, St Louis, MO) for urine with dye staining method was used for determination of urine protein concentration. The reference measurement instrument for urea, creati-
241
nine, glucose, protein and ketone concentration determinations was the Abbott Vision instrument. The Abbott Vision was not originally intended as a urinalysis system, but the standard procedures were adapted to serve with adequate precision. The estimated standard deviation for measurement of for urea, creatinine, glucose, protein and ketone in urine are 0.48 g/L, 0.03 g/L, 0.28 mmol/L, 0.01 g/L and 0.05 g/L, respectively.
4. Data analysis The first derivative of each spectrum was taken to enhance the separation and uniqueness of the spectral features of the individual analytes, and to eliminate baseline drift. The first-derivative spectra were calculated using a sliding spectral segment of 10 cm⫺1 with a gap of 2 cm⫺1. We determined that the first-derivative was preferred over higher order derivatives based on the following observations: [1] the baseline drift of our spectrometer was spectrally flat, and its first-derivative was vanishingly small, [2] the separation of the spectral features of the analytes of interest was sufficient for good regression using the firstderivative, [3] higher order derivatives, while increasing the separation and uniqueness of the relevant spectral features, increased the noise of the spectra to the point of diminishing the performance of the partial least squares (PLS) regression. The advantages of derivative spectroscopy including noise considerations have been reviewed in the literature [15–17]. Since the molar concentrations of urine analytes are relatively small compared to water’s molar concentration, the spectral signatures of the analytes of interest are very small compared to the spectral signature of water. To observe the spectral features of the analytes of interest, the water background must be subtracted from the urine spectra. Because each point of each spectrum is a large number, and because one large number must be subtracted from another large number to observe a small difference, both of the spectra must be determined precisely. For small concentrations of analyte, in some portions of the spectra, the signalto-noise ratio for determining this difference can approach 1. Figure 3 shows an example first-derivative spectra of glucose after water has been subtracted away. Below 2100 nm, the noise is of comparable size to the underlying spectral feature of interest. However, in this part of the spectrum, the spectral features, and their first derivative are slowly varying. Thus, it should be straight- forward to separate the spectral feature of interest from the rapidly varying noise. To separate the random noise from the underlying spectra, a sliding second order polynomial filter is applied. A second-order polynomial is fit to a 30 nm segment of the spectra, and a value computed by the fit is applied to the center of the 30 nm segment. The spectra is shifted by one spectral element (2 cm⫺1), and the previous step is repeated. A second order polynomial is chosen because it fits the 30
242
J.L. Pezzaniti et al. / Clinical Biochemistry 34 (2001) 239 –246
Fig. 3. Plot of first-derivative spectra of glucose at 10 g/L. Black smooth line shows smoothed first-derivative spectra of glucose. Gray choppy line shows unsmoothed first-derivative spectra of glucose.
nm wide underlying spectral features very well, and is a very poor fit to the rapidly fluctuating random noise. Figure 4 shows the residual noise, the noise subtracted from the original first-derivative spectra. This noise must be removed from the spectra before regression is done, to avoid the regression algorithm finding coincidental but spurious correlations between the spectra and the analyte concentrations.
5. Regression The near-infrared absorption of biologic materials are due to the overtone and combination bands of the molecular vibrations of C-H, O-H and N-H bonds, stretching vibrations, and O-H bending vibrations. These absorption bands are wide (tens of nanometers) and weak (a few percent of
Fig. 4. Residual noise after smoothing of glucose first-derivative spectra.
the absorption of water). When several biologic compounds are present with comparable concentrations in a matrix, the absorption bands overlap one another. At any given wavelength, many substances contribute to the measured absorption. To quantify a component of interest, absorption measurements must be made at wavelengths where the component absorbs and at additional wavelengths to compensate for interferants which spectrally overlap the component of interest. A variety of mathematical techniques, including classical least squares (CLS), partial least squares (PLS) and principle component analysis (PCA), have been applied to the problem of quantifying a compound of interest in the presence of interferants with similar spectral signature. PLS and PCA have been described [18]. The utility of CLS, PLS and PCA have been compared [19], and in-depth examples of applications of PLS have been described [20]. In the work presented in this paper, PLS was used to derive calibration equations which relate NIR spectra to the concentration of the biologic component of interest. In this method, actual samples are used so that in addition to calibrating to the signal from the analyte of interest, the calibration equation can compensate for interfering absorptions from other matrix components, shifts in analyte absorption band positions, and other sources of interference. In the calibration sample set, the concentrations of the analyte of interest should span the range over which the calibration will apply, and the interferants should be representative of those expected in the samples in which the calibration will be applied. Commercially available PLS software (MatLab Chemometrics, Math Works Inc., Natick, MA) was used to find optimal correlation between variations within the spectral data set and the corresponding analyte concentrations. We empirically determined the optimal wavelength segments for use in the PLS calibration to be between 1350 to 1800 and 2050 to 2375 nm. The region between 1800 nm and 2050 nm is excluded because this region is dominated by water absorption. Furthermore, because temperature changes in water cause the absorption peak to shift, this region is also dominated by temperature dependence. Figures 5 and 6 shows the first-derivative spectra of the five analytes of interest between 2100 to 2400. The region between 2000 and 2500 nm is the best for spectral analysis of organic molecules due to their strong absorption in this region. The spectral signatures of the five analytes are sufficiently unique to provide analytical specificity. Their relative concentration are however important. Table 1 gives the peak-to-peak amplitude of the most prominent spectral feature in the 2000 to 2500 nm range, the concentration of the analytes in urine, and the scaled signal contribution expected from the analyte in urine. This shows the large difference in the contribution of the signal from the various analytes. Urea is a factor of 10 larger than creatinine and at least 200 times larger than the other three analytes. Thus we would expect that the accuracy of NIR to measure these
J.L. Pezzaniti et al. / Clinical Biochemistry 34 (2001) 239 –246
243
Table 1 Summary of approximate peak-to-peak signal strengths of urea, creatinine, glucose ketone and protein in urine of healthy donors Mean Normal Peak-to-peak Peak-to-Peak spectral Concentration (g/L) Spectral feature feature for mean for 10 g/L Normal concentration Urea 16 Creatinine 1.1 Glucose 0.40 Protein 0.05 Ketone 0.05
Fig. 5. First-derivative spectra of urea, creatinine, and glucose, at concentrations of 10 g/L.
analytes in urine to follow the descending order of concentration in urine.
6. Calibration and model validation Choosing the appropriate number of PLS factors to use in the calibration equation is critical issue in PLS calibration. If too few factors are used, then the model will not adequately describe the system. If too many factors are used, the model will over-fit the system, and will not adequately predict concentrations for samples outside the model. One method for determining the appropriate number of PLS factors is cross-validation. In this method the data set is divided into samples used for calibration and samples used for model validation, say half for calibration and half for model validation. The ability of the PLS model generated from the calibration set is compared to the number of factors used in the calibration model. The standard error of prediction (SEP) is the rms of the difference between the
Fig. 6. First-derivative spectra of ketone and protein, at concentrations of 10 g/L.
0.007 0.008 0.001 0.005 0.003
8E—3 9E—4 4E—5 2E—5 1E—5
concentration of the model validation sample predicted by the model and the value determined by reference method. A similar metric, the Standard Error of Calibration (SEC) (the rms of the difference between the concentration of the calibration sample predicted by the model and the value determined by reference method) is also useful. SEC is a measure of how well the model describes the variability in the system, and the SEP is a measure of how much the model is overfitting the system. In the work reported here, the Standard Error of Prediction (SEP) and the correlation coefficient (r) were plotted against the number of PLS factors. The SEP values were observed to decrease continuously as more of the spectral variation was modeled. The point at which the model became overfit, the SEP would increase rapidly. We chose the number of factors by the location of this inflection point.
7. Results and discussion 7.1. Urea We determined that the optimal wavelengths for use in a PLS calibration and model validation was between 2050 nm and 2275 nm. Wavelengths between 1800 nm and 2050 nm were excluded because of the strong dependence of the water spectra on temperature. The spectral signature of urea is much larger than the components that make up the background matrix in urine. In fact a strong correlation between urea concentration and the derivative spectra at one wavelength can be found at 2137 nm. The correlation coefficient was r ⫽ 0.99 and the SEP ⫽ 1.12 g/L. The derivative was computed by subtracting two averaged 10 cm⫺1 segments 2 cm⫺1 apart. Shaw et al. [1] have shown that absorption values at three wavelengths (2152 nm primary wavelength, 1194 nm and 1724 nm for interference correction) is sufficient to obtain good prediction for urea, and SEP of about 1.0 g/L. A PLS calibration and model validation for Urea is shown in Figure 7 using the experimental protocol described above. The SEC, SEP and r values are given in Table 2. Although the spectral signature of urea is of comparable strength to that of creatinine, the SEC and SEP of urea is
244
J.L. Pezzaniti et al. / Clinical Biochemistry 34 (2001) 239 –246
Fig. 7. NIR calculated vs. reference lab values for urea calibration and model validation sets, using protocol 1.
much larger than that of creatinine (see Table 2). The majority of the discrepancy of performance is caused by the difference in accuracy between the reference assay for urea and creatinine. The urea reference assay has a 2.5% CV plus a 1% dilution error need for getting the sample concentration within the range of the Vision instrument. The mean concentration for the sample set is approximately 13 g/L, so the error is approximately 0.46 g/L. The precision of the creatinine reference assay is much lower, 0.7% plus 1% dilution error. The mean concentration of the sample set is 1.75 g/L. Thus, the reference error for creatinine is 0.03 g/L.
Fig. 8. NIR calculated vs. reference lab values for creatinine calibration and model validation sets, using protocol 1.
The spectral contribution of creatinine to the spectra of urine is smaller than that of urea, due to its lower concentration. The wavelength range that was used for PLS calibration model validation for creatinine was 1350 to 1800 nm, 2050 to 2375 nm. Although the spectral signature of creatinine is small in the wavelength range below 1800 nm, these additional wavelengths increased the accuracy of the validation slightly. The improvement in the SEP is from 0.21 to 0.17 g/L. The improvement is most likely due to the corrections for the interferences from other chemical components within the urine matrix. The accuracy for this measurement is shown in Figure 8.
the concentration of glucose in normal urine is small, ranging between 0.55 mmol/L/dL in a fasting urine to approximately 4.4 mmol/L after a meal for a normal “nondiabetic” person. Thus the spectral contribution of glucose to the urine matrix is very small. Accordingly, the SEP of this assay is relatively high. Unlike the high SEP for urea, this is not caused by error in the reference assay, but rather by the inherent difficulty in separating a spectral feature that is of comparable size to the background signal. Since glucose concentration in normal urine is usually lower than 2.8 mmol/L, additional glucose was added into approximately half of the samples to simulate values observed for diabetic patients. The distribution of glucose spanned 0 to approximately 44 mol/L. This was done to enhance the accuracy of the PLS calibration. A significant portion of the error in the measurement is caused by variation of the concentrations of interferences from sample to sample. Increasing the range of the measurement increases the signal relative to the background of the calibration, resulting in a more robust calibration. Figure 9 shows the performance of the spectral measurements against the reference assay. The clustering of values around various concentrations is the result of sample preparation decisions and the use of discrete glucose concentrations.
7.3. Glucose
7.4. Protein
The spectral features associated with glucose are small compared to urea and creatinine (see Table 1). In addition,
The total protein concentration in serum have been measured by NIR spectroscopic methods with an accuracy com-
7.2. Creatinine
Table 2 Summary of calibration and model validation performance using PLS regression for urea, creatinine, glucose, ketone and protein in urine using protocols 1 and 2
Protocol Factors SEC rcalibration SEP rvalidation Slope Intercept
Urea
Creatinine
Glucose
Protein
Protein
Ketone
Ketone
1 11 9.6 0.985 9.3 0.993 0.965 4.8
1 6 1.7 0.978 1.3 0.970 0.930 0.95
1 10 1.9 0.997 4.3 0.958 1.08 ⫺1.5
1 11 1.3 0.938 2.7 0.823 1.02 ⫺0.035
2 5 1.7 0.99 1.8 0.99 0.98 2.0
1 8 2.4 0.817 2.9 0.500 0.486 1.0
2 9 1.2 0.99 2.0 0.96 0.99 0.74
J.L. Pezzaniti et al. / Clinical Biochemistry 34 (2001) 239 –246
Fig. 9. NIR calculated vs. reference lab values for glucose calibration and model validation sets, using protocol 1.
parable to standard clinical chemistry methods [4]. Normal serum protein concentration is 2 to 9 g/L. However, protein concentration in normal urine, at about 0.05 g/L is as much as three orders of magnitude lower than the protein concentration in normal serum. Exercise can cause levels of urine protein to reach approximately 3000 mg/L for normal daily urine output of 800 to 1600 mL (Textbook of Clinical Chemistry pg 1266). Other disease conditions, such as Nephrotic syndrome, can produce urine protein concentrations of greater than 3 g/L. Measurement of protein levels that might be encountered in urine requires that the system accurately accounts for interferences and provides the necessary assay range. The following wavelengths were used for the PLS calibration for total urine protein: 1350 to 1800 and 2050 to 2375 nm.. Figure 10 shows the performance of the NIR measurements compared with the reference assay. Normal urine samples were spiked with BSA to provide a range of protein concentration in the calibration and test group samples that was between 0 and 1.50 g/L. For the protein concentration range studied, the SEP was 0.27 g/L. We carried out a second experiment with the goal of improving the accuracy of the method, protocol 2 described above. In this experiment, normal urines were spiked with BSA to obtain urine protein concentrations of 0, 0.02, 0.04, 0.06, 0.08, 0.10, 0.2, 0.4, 0.6, 0.8, 1.0 g/L. The uneven distribution of sample concentrations was chosen to emphasize performance at lower urine protein concentrations. The
Fig. 10. NIR calculated vs. reference lab values for protein calibration and model validation sets, using protocol 1.
245
Fig. 11. NIR calculated vs. reference lab values for protein calibration and model validation sets, using protocol 2.
wider range of sample protein concentration was chosen so that the spectral signal of protein was large compared to the background interference. The wider range enables a more accurate PLS calibration model to be developed by emphasizing in the calibration the wavelengths that most critically affect the determination of urine protein concentration. For these measurements 30 urine samples were used for calibration and 30 independent urine samples were used for model validation as described in protocol 2. Figure 11 shows the performance for the NIR against the reference assay. The SEP for this method is 0.18 g/L, producing the expected improvement in assay precision. While only BSA was used as a model protein in these studies, other proteins might be expected to respond similarly as the wavelengths used are overtones of fundamental peptide vibrations. If further studies prove the supposition proves correct, NIR protein concentration measurement method could have wide applicability in urinalysis and other clinical chemistry measurements. 7.5. Ketones Normal urine might be expected to contain less than 0.05 g/L of total ketones. The average composition of urinary ketones is expected to be 80% - hydroxybuterate, 17% acetoacetate, and 3% acetone. We measured - hydroxybuterate, the most prevalent urine ketone component. Because only normal urine samples were available for testing, we enhanced the range of concentrations of ketones by spiking in - hydroxybuterate to increase the range to 1.50 g/L. The standard deviation for the - hydroxybuterate reference assay was 0.05 g/dL. Two experiments were carried out to evaluate the performance of NIR measurements for ketones. The protocols were similar to the protein experiments described above. In the first protocol, sixty samples were used for calibration, and a second set of sixty samples was used for model validation. Figure 12 shows the performance of the NIR measurements against the reference assay. The range of ketone bodies was 0 to 1.50 g/L. The SEP for the ketone bodies was 0.29 g/L. A second experiment was conducted using protocol 2,
246
J.L. Pezzaniti et al. / Clinical Biochemistry 34 (2001) 239 –246
The accuracy and precision obtainable using NIR for urea and creatinine is comparable to that obtained with typical laboratory standards. Currently, the accuracy for glucose, ketone and protein is just sufficient to be a useful as a diagnostic screening assay. Further improvements in assay precision might be expected with improved reference assay precision, by employing other or additional wavelength ranges, and by instrument improvements that further reduce or cancel noise. Fig. 12. NIR calculated vs. reference lab values for ketone calibration and model validation sets, using protocol 1.
with the goal of improving the accuracy of the method. In this experiment, normal urines were spiked with -hydroxybuterate to obtain several different concentrations. The levels used were 0, 0.2, 0.4, 0.6, 0.8, 1.0, 2.0, 4.0, 6.0, 8.0 g/L. For this measurement 30 samples were used for calibration and 30 independent samples were used for model validation. Figure 13 shows the performance for the NIR against the reference assay. The SEP for this method is 0.2 g/L.
8. Summary Urinary urea, creatinine, glucose, ketone, and protein can be quantified using near-infrared spectroscopy. PLS models, with an optimal number of factors can quantitate urea, creatinine, glucose, ketone and protein in urine with SEP’s of 0.93 g/L, 0.13 g/L, 4.3 mmoL/L, 0.2 g/L and 0.18 g/L, respectively. The precision of the method improves significantly with calibration samples that include very high concentrations. The high analyte concentrations emphasize the spectral features of the analyte vs. the background. For example, the precision of urinary protein measurements using a sample concentration between 0 and 1.5 g/L for calibration was 0.27 g/L. Increasing the range of urinary protein concentration to 0 to 7.0 g/L decreased the SEP to 0.18 g/L. The ketone performance SEP was 0.29 g/L with calibration 0 to 1.50 g/L. Increasing the concentration range of calibration samples to 0 to 8.0 g/L reduced the SEP to 0.2 g/L.
Fig. 13. NIR calculated vs. reference lab values for ketone calibration and model validation sets, using protocol 2.
References [1] Shaw RA, Kotowich S, Mantsch HH, Leroux M. Quantitation of protein, creatinine, and urea in urine. Clin Biochem 1996;29:11–19. [2] Van Toorenberger AW. Assay of glucose in urine by near-infrared spectrophotometry. Clin Chem 1994;40:1788. [3] Hall JW, Pollard A. Near-infrared spectophotometry: A new dimension in clinical chemistry. Clin Chem 1992;38:1623–31. [4] Hall JW, Pollard A. Near-Infrared spectroscopic determination of serum total proteins, albumin, globulins, and urea. Clin Biochem 1993;26:483–90. [5] Peuchant E, Salles C, Jensen R. Value of a spectroscopic “fecalogram” in determining the etiology of steatorrhea. Clin Chem 1998; 34:5– 8. [6] Koumantakis G, Radcliff FJ. Estimating fat in feces by near-infrared reflectance spectroscopy. Clin Chem 1987;34:502– 6. [7] Pan S, Chung H, Arnold M. Near-infrared spectroscopic measurement of physiological glucose levels in variable matrices of protein and triglycerides. Anal Chem 1996;68:1124 –34. [8] Heise HM, Marbach R, Janatsch G, Kruse-Jarres JD. Multivariate determination of glucose in whole blood by attenuated total reflection infrared scpectroscopy. Anal Chem 1989;61:2009 –15. [9] Voswinckel P. A marvel of colors, and ingredients: The story of urine test strips. Kidney International 1994;46:S3–S7. [10] Free AH, Free HM. Urinalysis: Its proper role in the physician’s office. Office Practice of Laboratory Medicine 1986;6:253– 66. [11] Bakhshandeh S, Morita Y. Comparison of urinary concentration tests: osmolarity, specific gravity, and refractive index. Michigan Medicine 1975;74:399 – 403. [12] Low PS, Tay JSH. Urine - osmolality, refrative indes, and specific gravity. The Journal of the Singapore Paedoatric Society 1978;20: 37– 42. [13] Watson ID. Analysis of commonly abused drugs in urine a selected threshold (cutoff) concentrations. Clini Chemy 1992;38:441. [14] George S, Briathwaite RA. An investigation into the extent of possbile dilution of specimens received for urinary drugs of abuse screening. Addiciton 1995;90:967–70. [15] O’Havier TC, Begley T. Signal-to-noise ratio in higher order derivative spectrometry. Anal Chem 1981;53:1876 – 8. [16] Levillain P, Pompeydie D. Derivative spectrophotometry: principles, advantages and limitations, applications. Analysis 1986;14:1–20. [17] Fell AF. Biomedical applications of derivative spectroscopy. Trends Anal Chem 1983;2:63– 6. [18] Martens H, Naes T. Multivariate Calibration. New York: John Wiley & Sons, 1989. [19] Thomas EV, Haaland DM. Comparison of multivariate calibration methods for quantitative spectral analysis. Anal Chem 1990;62:1091–9. [20] Lin CT. Applications of partial least squares to spectral data for quantitative determinations of chemicals in a mixture. Drug Information Journal 1990;24:671– 89.
Preliminary investigation of near-infrared spectroscopic measurements of urea, creatinine, glucose, protein, and ketone in urine J. Larry Pezzaniti, Tzyy-Wen Jeng, Larry McDowell, Gary M. Oosta* Abbott Laboratories, Diagnostic Division, Abbott Park, IL 60064, USA Received 14 November 2000; received in revised 14 February 2001; accepted 16 February 2001
Abstract Objective: We investigated the use of near-infrared spectroscopy as an analytical tool to quantify concentrations of urea, creatinine, glucose, ketone, and protein in urine. Design and Methods: FT-IR spectroscopy in conjunction with a polynomial based spectral smoothing method was applied to urine specimens. A partial factorial experimental design was employed to collect spectra using normal and spiked urine samples. Results: Our results show that the spectral signatures of urea, creatinine, glucose, ketone, and protein in the 1350 to 1800 nm and 2050 to 2375 nm range are sufficiently strong and unique for accurate measurements. Conclusions: The accuracy of near infrared for quantifying concentrations of urea and creatinine is only slightly less than our selected reference methods. Glucose, ketone and protein are sufficiently accurate to be useful as a screening tool for wellness. The method successfully accounts for biologic matrix variation. The advantages of near-infrared analysis are (1) no reagents, (2) ease of sample preparation, (3) speed, and (4) the ability to quantify multiple analytes with one spectra. © 2001 The Canadian Society of Clinical Chemists. All rights reserved. Keywords: Urinalysis; Urine; Spectroscopy; Infrared; Near-infrared; Visible; Fourier transform
1. Introduction For clinical laboratory measurements, quantitative methods must be suitably accurate and precise over the expected range of values required. In addition, it is often desirable that the method be inexpensive, reliable, rapid and easily automated. Near-infrared spectroscopy has the potential to satisfy these criteria. It needs no reagents, little or no sample preparation, it is rapid and nondestructive, and is suitable for complex matrices. Near-infrared spectroscopy has been applied to measuring urine composition [1,2], serum composition [3,4], fecal composition [5,6], glucose in whole blood [7] and complex matrices [8]. The paper explores the feasibility of the use of near-infrared analysis for measuring five compounds of interest for urinalysis screening testing. Urine contains a wide variety of substances. Urinalysis involves measuring critical components in a sample of urine to identify previously undetected diseases or medical con* Corresponding author. Tel.: ⫹1-847-937-2553; fax: ⫹1-847-9386929. E-mail address: [email protected] (G.M. Oosta).
ditions, or urinalysis may be used to determine if regulated substances (e.g., drugs) are being abused. All of these analytes represent breakdown products of metabolism from various organ systems. The pattern of excretion is indicative of various disease states. The history and utility of urinalysis have been reviewed [9,10]. In current urinalysis systems, such as those provided by Bayer and Boehringer Mannheim, the analytes measured include glucose, bilirubin, blood (or hemoglobin), protein, urobilinogen, nitrites, leukocytes, specific gravity, and pH. In urine the major ketone components are 3-hydroxybutyrate (80%), acetoacetic acid (17%) and acetone (3%), but only the acetoacetic acid is determined by the current test systems. Refractive index may be substituted for specific gravity [11,12]. In some cases, measurement of creatinine is suggested, but is not provided by Bayer’s or Boehringer Mannheim’s urinalysis systems [13,14]. The majority of urinalysis testing is accomplished by means of dip and read strip technology supplied by Bayer and Boehringer Mannheim. Strip technology is well understood, and suffers from a number of limitations. Readings must be properly timed to obtain accurate results. Urine
0009-9120/01/$ – see front matter © 2001 The Canadian Society of Clinical Chemists. All rights reserved. PII: S 0 0 0 9 - 9 1 2 0 ( 0 1 ) 0 0 1 9 8 - 9
240
J.L. Pezzaniti et al. / Clinical Biochemistry 34 (2001) 239 –246
samples must be well mixed and at room temperature. Strips are sensitive to light and humidity, and must be stored and handled properly. Quantitative results are difficult to obtain. Interfering substances can cause incorrect readings. This work demonstrates the utility of near-infrared analysis for measurement of glucose, ketone and protein (currently done with test strips), and urea and creatinine. The objective was to determine the accuracy of near-IR spectroscopy for determining concentrations of these analytes in urine.
2. Experimental protocol and instrumentation Two experimental protocols, referred to as protocols 1 and 2, were used to prepare the urine samples and collect the NIR data. For measurements involving urea, creatinine, glucose, experimental protocol 1 was employed. For protein and ketone, both experimental protocols 1 and 2 were used. Protocol 1 was designed to produce samples of varying analyte concentration and to provide a variable matrix from which an accurate partial least squares (PLS) model could be derived and to provide an independent set of samples that could be used to check the accuracy of the PLS model. Protocol 1 consisted of the following steps. A. First morning urine specimens were collected from sixty different healthy donors. B. The samples were filtered with a 0.45 micron pore size cellulose acetate filter (Nalge Nunc International, Inc.). C. Each specimen was divided into two portions to make two identical sample sets, samples 1 to 60 and samples 61 to 120. D. Samples 1 to 60 were used without change. E. Additional analytes were added to samples 61 to 120 according to a randomization protocol. In some cases, two or more analytes were added to the same sample. The randomization process provided samples that gave the high and low values needed for a calibration sample set and the high and low test sample concentrations that were needed for a validation sample set. In protocol 1, the final analyte concentrations were within the range of concentrations that might be expected in a clinical laboratory. Even numbered samples (1–120) were used as a validation set to check the accuracy of the PLS model that was developed. F. Odd numbered samples (1–120) were used as the calibration set for the development of the PLS model. Figure 1 shows the distributions for the concentrations of the analytes in samples 1 to 120 used for protocol. Protocol 2, while similar to protocol 1, incorporated the following changes. 1. Thirty first morning urine samples were used without prefiltration 2. When analytes were added as described in step E, the resulting concentrations were substantially greater than might be expected in a clinical laboratory
Fig. 1. Histogram showing the distribution of concentrations of urea, creatinine, glucose, protein and ketone in calibration and model validation samples, for protocol 1.
Figure 2 shows the distributions for the concentrations of the analytes in samples used for protocol 2. A sample handling system was used to deliver the samples to the cuvette that was held at a constant temperature of approximately 25.0 C. An automated Gilson 223 Sample Changer (Gilson Inc.) was used in conjunction with a constant temperature sample delivery system. The sample delivery system consisted of a 100 cm of 0.076 cm inside diameter aluminum tube submerged in a water bath maintained at 25.0 C ⫾ 0.01. The sample was be pumped through the aluminum tube with a peristaltic pump such that the sample would arrive at the cuvette at the desired temperature. The cuvette was held in thermal contact with a water-jacketed cuvette holder also maintained at 25.0 C ⫾ 0.01. The standard deviation of the measured sample to sample temperature was less than 0.02 C, as measured with a k-type thermocouple in contact with the urine. Carry-over
Fig. 2. Histrogram showing the distribution of protein and ketone in sample sets spiked with analyte beyond the normal range for healthy urines, for protocol 2.
J.L. Pezzaniti et al. / Clinical Biochemistry 34 (2001) 239 –246
of one sample to the next was reduced to less than 0.05% by a wash protocol between samples. Near-IR spectra were collected with an Equinox 55/S FTIR spectrometer (Bruker Inc.). The settings of the spectrometer where optimized for NIR measurements by using its 100W tungsten-halogen lamp, its Indium Galium Arsenide thermoelectrically cooled detector, and its quartz beam splitter. The spectral range of the measurements was 1100 to 2500 nm. The cuvette was a quartz flow through cell with a pathlength of 1 mm. To achieve a satisfactory signalto-noise ratio, 32 spectra were averaged at a spectral resolution of 2 cm⫺1
3. Sample preparation and reference methods Fresh urine samples from normal, healthy volunteers were collected for initial calibration and model validation testing. Urea (Aldrich Chemical Co., Milwaukee, WI) powder was weighed out to the nearest 0.1 mg, and dissolved into 10 ml of sample so that the sample set’s concentrations were of sufficient dynamic range for calibration. Urinary Urea Nitrogen concentration of both native urine and modified urine was determined by the Vision Instrument (Abbott) after predilution with distilled water. Appropriate dilution factor (usually between 51–26 fold) was used so that the diluted sample would contain urea within the normal range of the Vision Urea Nitrogen Assay which had been developed and optimized for serum. Creatinine hydrochloride (Sigma, St. Louis, MO) was weighed out to the nearest 0.1 mg, and dissolved into 10 ml of sample so that the sample set’s concentrations were of sufficient dynamic range for calibration. Urinary creatinine concentration of both native urine and modified urine was determined by the Vision Instrument (Abbott) after predilution with distilled water. Appropriate dilution factor (usually between 33–26 fold) was used so that the diluted sample would contain creatinine within the functional range of the Vision Creatinine Assay which had been developed and optimized for serum. 3-Hydroxybutyric Acid (Aldrich Chemical Co., Milwaukee, WI) powder was weighed out to the nearest 0.1 mg, and dissolved into 10 ml of sample so that the sample set’s concentrations were of sufficient dynamic range for calibration. An enzymatic method for assaying -hydroxybutyrate from Sigma Diagnostics (St. Louis, MO) was used to measure the concentration of ketone in both the native and spiked urine samples. BSA Stock solution was prepared by dissolving 40 g of bovine serum albumin (BSA) powder (Fraction V, Sigma Chemical Co., St. Louis, MO) per liter. It was stored at 4°C. The Micro-protein Assay (Sigma Diagnostics, St Louis, MO) for urine with dye staining method was used for determination of urine protein concentration. The reference measurement instrument for urea, creati-
241
nine, glucose, protein and ketone concentration determinations was the Abbott Vision instrument. The Abbott Vision was not originally intended as a urinalysis system, but the standard procedures were adapted to serve with adequate precision. The estimated standard deviation for measurement of for urea, creatinine, glucose, protein and ketone in urine are 0.48 g/L, 0.03 g/L, 0.28 mmol/L, 0.01 g/L and 0.05 g/L, respectively.
4. Data analysis The first derivative of each spectrum was taken to enhance the separation and uniqueness of the spectral features of the individual analytes, and to eliminate baseline drift. The first-derivative spectra were calculated using a sliding spectral segment of 10 cm⫺1 with a gap of 2 cm⫺1. We determined that the first-derivative was preferred over higher order derivatives based on the following observations: [1] the baseline drift of our spectrometer was spectrally flat, and its first-derivative was vanishingly small, [2] the separation of the spectral features of the analytes of interest was sufficient for good regression using the firstderivative, [3] higher order derivatives, while increasing the separation and uniqueness of the relevant spectral features, increased the noise of the spectra to the point of diminishing the performance of the partial least squares (PLS) regression. The advantages of derivative spectroscopy including noise considerations have been reviewed in the literature [15–17]. Since the molar concentrations of urine analytes are relatively small compared to water’s molar concentration, the spectral signatures of the analytes of interest are very small compared to the spectral signature of water. To observe the spectral features of the analytes of interest, the water background must be subtracted from the urine spectra. Because each point of each spectrum is a large number, and because one large number must be subtracted from another large number to observe a small difference, both of the spectra must be determined precisely. For small concentrations of analyte, in some portions of the spectra, the signalto-noise ratio for determining this difference can approach 1. Figure 3 shows an example first-derivative spectra of glucose after water has been subtracted away. Below 2100 nm, the noise is of comparable size to the underlying spectral feature of interest. However, in this part of the spectrum, the spectral features, and their first derivative are slowly varying. Thus, it should be straight- forward to separate the spectral feature of interest from the rapidly varying noise. To separate the random noise from the underlying spectra, a sliding second order polynomial filter is applied. A second-order polynomial is fit to a 30 nm segment of the spectra, and a value computed by the fit is applied to the center of the 30 nm segment. The spectra is shifted by one spectral element (2 cm⫺1), and the previous step is repeated. A second order polynomial is chosen because it fits the 30
242
J.L. Pezzaniti et al. / Clinical Biochemistry 34 (2001) 239 –246
Fig. 3. Plot of first-derivative spectra of glucose at 10 g/L. Black smooth line shows smoothed first-derivative spectra of glucose. Gray choppy line shows unsmoothed first-derivative spectra of glucose.
nm wide underlying spectral features very well, and is a very poor fit to the rapidly fluctuating random noise. Figure 4 shows the residual noise, the noise subtracted from the original first-derivative spectra. This noise must be removed from the spectra before regression is done, to avoid the regression algorithm finding coincidental but spurious correlations between the spectra and the analyte concentrations.
5. Regression The near-infrared absorption of biologic materials are due to the overtone and combination bands of the molecular vibrations of C-H, O-H and N-H bonds, stretching vibrations, and O-H bending vibrations. These absorption bands are wide (tens of nanometers) and weak (a few percent of
Fig. 4. Residual noise after smoothing of glucose first-derivative spectra.
the absorption of water). When several biologic compounds are present with comparable concentrations in a matrix, the absorption bands overlap one another. At any given wavelength, many substances contribute to the measured absorption. To quantify a component of interest, absorption measurements must be made at wavelengths where the component absorbs and at additional wavelengths to compensate for interferants which spectrally overlap the component of interest. A variety of mathematical techniques, including classical least squares (CLS), partial least squares (PLS) and principle component analysis (PCA), have been applied to the problem of quantifying a compound of interest in the presence of interferants with similar spectral signature. PLS and PCA have been described [18]. The utility of CLS, PLS and PCA have been compared [19], and in-depth examples of applications of PLS have been described [20]. In the work presented in this paper, PLS was used to derive calibration equations which relate NIR spectra to the concentration of the biologic component of interest. In this method, actual samples are used so that in addition to calibrating to the signal from the analyte of interest, the calibration equation can compensate for interfering absorptions from other matrix components, shifts in analyte absorption band positions, and other sources of interference. In the calibration sample set, the concentrations of the analyte of interest should span the range over which the calibration will apply, and the interferants should be representative of those expected in the samples in which the calibration will be applied. Commercially available PLS software (MatLab Chemometrics, Math Works Inc., Natick, MA) was used to find optimal correlation between variations within the spectral data set and the corresponding analyte concentrations. We empirically determined the optimal wavelength segments for use in the PLS calibration to be between 1350 to 1800 and 2050 to 2375 nm. The region between 1800 nm and 2050 nm is excluded because this region is dominated by water absorption. Furthermore, because temperature changes in water cause the absorption peak to shift, this region is also dominated by temperature dependence. Figures 5 and 6 shows the first-derivative spectra of the five analytes of interest between 2100 to 2400. The region between 2000 and 2500 nm is the best for spectral analysis of organic molecules due to their strong absorption in this region. The spectral signatures of the five analytes are sufficiently unique to provide analytical specificity. Their relative concentration are however important. Table 1 gives the peak-to-peak amplitude of the most prominent spectral feature in the 2000 to 2500 nm range, the concentration of the analytes in urine, and the scaled signal contribution expected from the analyte in urine. This shows the large difference in the contribution of the signal from the various analytes. Urea is a factor of 10 larger than creatinine and at least 200 times larger than the other three analytes. Thus we would expect that the accuracy of NIR to measure these
J.L. Pezzaniti et al. / Clinical Biochemistry 34 (2001) 239 –246
243
Table 1 Summary of approximate peak-to-peak signal strengths of urea, creatinine, glucose ketone and protein in urine of healthy donors Mean Normal Peak-to-peak Peak-to-Peak spectral Concentration (g/L) Spectral feature feature for mean for 10 g/L Normal concentration Urea 16 Creatinine 1.1 Glucose 0.40 Protein 0.05 Ketone 0.05
Fig. 5. First-derivative spectra of urea, creatinine, and glucose, at concentrations of 10 g/L.
analytes in urine to follow the descending order of concentration in urine.
6. Calibration and model validation Choosing the appropriate number of PLS factors to use in the calibration equation is critical issue in PLS calibration. If too few factors are used, then the model will not adequately describe the system. If too many factors are used, the model will over-fit the system, and will not adequately predict concentrations for samples outside the model. One method for determining the appropriate number of PLS factors is cross-validation. In this method the data set is divided into samples used for calibration and samples used for model validation, say half for calibration and half for model validation. The ability of the PLS model generated from the calibration set is compared to the number of factors used in the calibration model. The standard error of prediction (SEP) is the rms of the difference between the
Fig. 6. First-derivative spectra of ketone and protein, at concentrations of 10 g/L.
0.007 0.008 0.001 0.005 0.003
8E—3 9E—4 4E—5 2E—5 1E—5
concentration of the model validation sample predicted by the model and the value determined by reference method. A similar metric, the Standard Error of Calibration (SEC) (the rms of the difference between the concentration of the calibration sample predicted by the model and the value determined by reference method) is also useful. SEC is a measure of how well the model describes the variability in the system, and the SEP is a measure of how much the model is overfitting the system. In the work reported here, the Standard Error of Prediction (SEP) and the correlation coefficient (r) were plotted against the number of PLS factors. The SEP values were observed to decrease continuously as more of the spectral variation was modeled. The point at which the model became overfit, the SEP would increase rapidly. We chose the number of factors by the location of this inflection point.
7. Results and discussion 7.1. Urea We determined that the optimal wavelengths for use in a PLS calibration and model validation was between 2050 nm and 2275 nm. Wavelengths between 1800 nm and 2050 nm were excluded because of the strong dependence of the water spectra on temperature. The spectral signature of urea is much larger than the components that make up the background matrix in urine. In fact a strong correlation between urea concentration and the derivative spectra at one wavelength can be found at 2137 nm. The correlation coefficient was r ⫽ 0.99 and the SEP ⫽ 1.12 g/L. The derivative was computed by subtracting two averaged 10 cm⫺1 segments 2 cm⫺1 apart. Shaw et al. [1] have shown that absorption values at three wavelengths (2152 nm primary wavelength, 1194 nm and 1724 nm for interference correction) is sufficient to obtain good prediction for urea, and SEP of about 1.0 g/L. A PLS calibration and model validation for Urea is shown in Figure 7 using the experimental protocol described above. The SEC, SEP and r values are given in Table 2. Although the spectral signature of urea is of comparable strength to that of creatinine, the SEC and SEP of urea is
244
J.L. Pezzaniti et al. / Clinical Biochemistry 34 (2001) 239 –246
Fig. 7. NIR calculated vs. reference lab values for urea calibration and model validation sets, using protocol 1.
much larger than that of creatinine (see Table 2). The majority of the discrepancy of performance is caused by the difference in accuracy between the reference assay for urea and creatinine. The urea reference assay has a 2.5% CV plus a 1% dilution error need for getting the sample concentration within the range of the Vision instrument. The mean concentration for the sample set is approximately 13 g/L, so the error is approximately 0.46 g/L. The precision of the creatinine reference assay is much lower, 0.7% plus 1% dilution error. The mean concentration of the sample set is 1.75 g/L. Thus, the reference error for creatinine is 0.03 g/L.
Fig. 8. NIR calculated vs. reference lab values for creatinine calibration and model validation sets, using protocol 1.
The spectral contribution of creatinine to the spectra of urine is smaller than that of urea, due to its lower concentration. The wavelength range that was used for PLS calibration model validation for creatinine was 1350 to 1800 nm, 2050 to 2375 nm. Although the spectral signature of creatinine is small in the wavelength range below 1800 nm, these additional wavelengths increased the accuracy of the validation slightly. The improvement in the SEP is from 0.21 to 0.17 g/L. The improvement is most likely due to the corrections for the interferences from other chemical components within the urine matrix. The accuracy for this measurement is shown in Figure 8.
the concentration of glucose in normal urine is small, ranging between 0.55 mmol/L/dL in a fasting urine to approximately 4.4 mmol/L after a meal for a normal “nondiabetic” person. Thus the spectral contribution of glucose to the urine matrix is very small. Accordingly, the SEP of this assay is relatively high. Unlike the high SEP for urea, this is not caused by error in the reference assay, but rather by the inherent difficulty in separating a spectral feature that is of comparable size to the background signal. Since glucose concentration in normal urine is usually lower than 2.8 mmol/L, additional glucose was added into approximately half of the samples to simulate values observed for diabetic patients. The distribution of glucose spanned 0 to approximately 44 mol/L. This was done to enhance the accuracy of the PLS calibration. A significant portion of the error in the measurement is caused by variation of the concentrations of interferences from sample to sample. Increasing the range of the measurement increases the signal relative to the background of the calibration, resulting in a more robust calibration. Figure 9 shows the performance of the spectral measurements against the reference assay. The clustering of values around various concentrations is the result of sample preparation decisions and the use of discrete glucose concentrations.
7.3. Glucose
7.4. Protein
The spectral features associated with glucose are small compared to urea and creatinine (see Table 1). In addition,
The total protein concentration in serum have been measured by NIR spectroscopic methods with an accuracy com-
7.2. Creatinine
Table 2 Summary of calibration and model validation performance using PLS regression for urea, creatinine, glucose, ketone and protein in urine using protocols 1 and 2
Protocol Factors SEC rcalibration SEP rvalidation Slope Intercept
Urea
Creatinine
Glucose
Protein
Protein
Ketone
Ketone
1 11 9.6 0.985 9.3 0.993 0.965 4.8
1 6 1.7 0.978 1.3 0.970 0.930 0.95
1 10 1.9 0.997 4.3 0.958 1.08 ⫺1.5
1 11 1.3 0.938 2.7 0.823 1.02 ⫺0.035
2 5 1.7 0.99 1.8 0.99 0.98 2.0
1 8 2.4 0.817 2.9 0.500 0.486 1.0
2 9 1.2 0.99 2.0 0.96 0.99 0.74
J.L. Pezzaniti et al. / Clinical Biochemistry 34 (2001) 239 –246
Fig. 9. NIR calculated vs. reference lab values for glucose calibration and model validation sets, using protocol 1.
parable to standard clinical chemistry methods [4]. Normal serum protein concentration is 2 to 9 g/L. However, protein concentration in normal urine, at about 0.05 g/L is as much as three orders of magnitude lower than the protein concentration in normal serum. Exercise can cause levels of urine protein to reach approximately 3000 mg/L for normal daily urine output of 800 to 1600 mL (Textbook of Clinical Chemistry pg 1266). Other disease conditions, such as Nephrotic syndrome, can produce urine protein concentrations of greater than 3 g/L. Measurement of protein levels that might be encountered in urine requires that the system accurately accounts for interferences and provides the necessary assay range. The following wavelengths were used for the PLS calibration for total urine protein: 1350 to 1800 and 2050 to 2375 nm.. Figure 10 shows the performance of the NIR measurements compared with the reference assay. Normal urine samples were spiked with BSA to provide a range of protein concentration in the calibration and test group samples that was between 0 and 1.50 g/L. For the protein concentration range studied, the SEP was 0.27 g/L. We carried out a second experiment with the goal of improving the accuracy of the method, protocol 2 described above. In this experiment, normal urines were spiked with BSA to obtain urine protein concentrations of 0, 0.02, 0.04, 0.06, 0.08, 0.10, 0.2, 0.4, 0.6, 0.8, 1.0 g/L. The uneven distribution of sample concentrations was chosen to emphasize performance at lower urine protein concentrations. The
Fig. 10. NIR calculated vs. reference lab values for protein calibration and model validation sets, using protocol 1.
245
Fig. 11. NIR calculated vs. reference lab values for protein calibration and model validation sets, using protocol 2.
wider range of sample protein concentration was chosen so that the spectral signal of protein was large compared to the background interference. The wider range enables a more accurate PLS calibration model to be developed by emphasizing in the calibration the wavelengths that most critically affect the determination of urine protein concentration. For these measurements 30 urine samples were used for calibration and 30 independent urine samples were used for model validation as described in protocol 2. Figure 11 shows the performance for the NIR against the reference assay. The SEP for this method is 0.18 g/L, producing the expected improvement in assay precision. While only BSA was used as a model protein in these studies, other proteins might be expected to respond similarly as the wavelengths used are overtones of fundamental peptide vibrations. If further studies prove the supposition proves correct, NIR protein concentration measurement method could have wide applicability in urinalysis and other clinical chemistry measurements. 7.5. Ketones Normal urine might be expected to contain less than 0.05 g/L of total ketones. The average composition of urinary ketones is expected to be 80% - hydroxybuterate, 17% acetoacetate, and 3% acetone. We measured - hydroxybuterate, the most prevalent urine ketone component. Because only normal urine samples were available for testing, we enhanced the range of concentrations of ketones by spiking in - hydroxybuterate to increase the range to 1.50 g/L. The standard deviation for the - hydroxybuterate reference assay was 0.05 g/dL. Two experiments were carried out to evaluate the performance of NIR measurements for ketones. The protocols were similar to the protein experiments described above. In the first protocol, sixty samples were used for calibration, and a second set of sixty samples was used for model validation. Figure 12 shows the performance of the NIR measurements against the reference assay. The range of ketone bodies was 0 to 1.50 g/L. The SEP for the ketone bodies was 0.29 g/L. A second experiment was conducted using protocol 2,
246
J.L. Pezzaniti et al. / Clinical Biochemistry 34 (2001) 239 –246
The accuracy and precision obtainable using NIR for urea and creatinine is comparable to that obtained with typical laboratory standards. Currently, the accuracy for glucose, ketone and protein is just sufficient to be a useful as a diagnostic screening assay. Further improvements in assay precision might be expected with improved reference assay precision, by employing other or additional wavelength ranges, and by instrument improvements that further reduce or cancel noise. Fig. 12. NIR calculated vs. reference lab values for ketone calibration and model validation sets, using protocol 1.
with the goal of improving the accuracy of the method. In this experiment, normal urines were spiked with -hydroxybuterate to obtain several different concentrations. The levels used were 0, 0.2, 0.4, 0.6, 0.8, 1.0, 2.0, 4.0, 6.0, 8.0 g/L. For this measurement 30 samples were used for calibration and 30 independent samples were used for model validation. Figure 13 shows the performance for the NIR against the reference assay. The SEP for this method is 0.2 g/L.
8. Summary Urinary urea, creatinine, glucose, ketone, and protein can be quantified using near-infrared spectroscopy. PLS models, with an optimal number of factors can quantitate urea, creatinine, glucose, ketone and protein in urine with SEP’s of 0.93 g/L, 0.13 g/L, 4.3 mmoL/L, 0.2 g/L and 0.18 g/L, respectively. The precision of the method improves significantly with calibration samples that include very high concentrations. The high analyte concentrations emphasize the spectral features of the analyte vs. the background. For example, the precision of urinary protein measurements using a sample concentration between 0 and 1.5 g/L for calibration was 0.27 g/L. Increasing the range of urinary protein concentration to 0 to 7.0 g/L decreased the SEP to 0.18 g/L. The ketone performance SEP was 0.29 g/L with calibration 0 to 1.50 g/L. Increasing the concentration range of calibration samples to 0 to 8.0 g/L reduced the SEP to 0.2 g/L.
Fig. 13. NIR calculated vs. reference lab values for ketone calibration and model validation sets, using protocol 2.
References [1] Shaw RA, Kotowich S, Mantsch HH, Leroux M. Quantitation of protein, creatinine, and urea in urine. Clin Biochem 1996;29:11–19. [2] Van Toorenberger AW. Assay of glucose in urine by near-infrared spectrophotometry. Clin Chem 1994;40:1788. [3] Hall JW, Pollard A. Near-infrared spectophotometry: A new dimension in clinical chemistry. Clin Chem 1992;38:1623–31. [4] Hall JW, Pollard A. Near-Infrared spectroscopic determination of serum total proteins, albumin, globulins, and urea. Clin Biochem 1993;26:483–90. [5] Peuchant E, Salles C, Jensen R. Value of a spectroscopic “fecalogram” in determining the etiology of steatorrhea. Clin Chem 1998; 34:5– 8. [6] Koumantakis G, Radcliff FJ. Estimating fat in feces by near-infrared reflectance spectroscopy. Clin Chem 1987;34:502– 6. [7] Pan S, Chung H, Arnold M. Near-infrared spectroscopic measurement of physiological glucose levels in variable matrices of protein and triglycerides. Anal Chem 1996;68:1124 –34. [8] Heise HM, Marbach R, Janatsch G, Kruse-Jarres JD. Multivariate determination of glucose in whole blood by attenuated total reflection infrared scpectroscopy. Anal Chem 1989;61:2009 –15. [9] Voswinckel P. A marvel of colors, and ingredients: The story of urine test strips. Kidney International 1994;46:S3–S7. [10] Free AH, Free HM. Urinalysis: Its proper role in the physician’s office. Office Practice of Laboratory Medicine 1986;6:253– 66. [11] Bakhshandeh S, Morita Y. Comparison of urinary concentration tests: osmolarity, specific gravity, and refractive index. Michigan Medicine 1975;74:399 – 403. [12] Low PS, Tay JSH. Urine - osmolality, refrative indes, and specific gravity. The Journal of the Singapore Paedoatric Society 1978;20: 37– 42. [13] Watson ID. Analysis of commonly abused drugs in urine a selected threshold (cutoff) concentrations. Clini Chemy 1992;38:441. [14] George S, Briathwaite RA. An investigation into the extent of possbile dilution of specimens received for urinary drugs of abuse screening. Addiciton 1995;90:967–70. [15] O’Havier TC, Begley T. Signal-to-noise ratio in higher order derivative spectrometry. Anal Chem 1981;53:1876 – 8. [16] Levillain P, Pompeydie D. Derivative spectrophotometry: principles, advantages and limitations, applications. Analysis 1986;14:1–20. [17] Fell AF. Biomedical applications of derivative spectroscopy. Trends Anal Chem 1983;2:63– 6. [18] Martens H, Naes T. Multivariate Calibration. New York: John Wiley & Sons, 1989. [19] Thomas EV, Haaland DM. Comparison of multivariate calibration methods for quantitative spectral analysis. Anal Chem 1990;62:1091–9. [20] Lin CT. Applications of partial least squares to spectral data for quantitative determinations of chemicals in a mixture. Drug Information Journal 1990;24:671– 89.