| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Articles |
1
Department of Pathology and Laboratory Medicine, University Hospital Groningen (UHG), PO Box 30001, 9700 RB Groningen, The Netherlands.
2
Department of Clinical Chemistry and Hematology (CKCHL), St. Elisabeth Hospital, PO Box 10111, 5000 JC Tilburg, The Netherlands.
aAuthor for correspondence. Fax 31-13-5352390; e-mail h.m.j.goldschmidt{at}ckchl-mb.nl.
 |
Abstract |
|---|
 Top Abstract IntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
Methods: The library was prepared from 25 pure components and 236 binary and ternary mixtures of the 8 most commonly occurring components. The ANN was trained and validated with 248 similar mixtures and tested with 92 patient samples, respectively.
Results: The optimum ANN model yielded root mean square errors of 1.5% and 2.3% for the training and validation sets, respectively. Fourteen simple expert rules were added to correct systematic network inaccuracies. Results of 92 consecutive patient samples were compared with those of a FT-IR method with KBr tablets, based on an initial computerized library search followed by visual inspection. The bias was significantly different from zero for brushite (-0.8%) and the concomitantly occurring whewellite (-2.8%) and weddellite (3.8%), but not for ammonium hydrogen urate (-0.1%), carbonate apatite (0.5%), cystine (0.0%), struvite (0.4%), and uric acid (-0.1%). The 95% level of agreement of all results was 9%.
Conclusions: The new Golden Gate method is superior because of its smaller sample size, user-friendliness, robustness, and speed. Expert knowledge for spectral interpretation is minimized by the combination of a library search and ANN prediction, but visual inspection remains necessary.
|
Introduction |
|---|
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
Fourier transform IR (FT-IR) spectroscopy has become a standard technique for analysis of urinary calculi. FT-IR makes use of a diversity of sample holders, such as photoacoustic detection (6), diffuse reflectance FT-IR (DRIFT) (7), and KBr tablet transmission (8)(9). For the routine visual interpretation of urinary calculus IR spectra, Hesse et al. (10) have issued an atlas with IR spectra from pure urinary calculus components and their mixtures, all embedded in KBr tablets. Another, less time-consuming option is computerized analyses based on a library search [e.g., SEARCH (11) and LITHOS (2)], expert rules [CIRCOM (12) and STONES (9)], or other chemometric techniques such as partial least-squares (PLS) regression (13) and artificial neural networks (ANNs) (14).
IR spectroscopy using KBr tablets is the current method for analysis of urinary calculus composition in our laboratory. The preparation of KBr tablets is time-consuming and often hampered by pellet breakage. To overcome these drawbacks, we developed a new IR method, using a Golden Gate Single Refection Diamond Attenuated Total Reflection (ATR) device. This method makes use of authentic sample material without any sample pretreatment. The results of the Golden Gate assay were quantified by a program dedicated for the prediction of the outcome of urinary calculus composition analyses. The new method was validated by comparing the results with those obtained by the IR assay with KBr tablets. The quantitative results from this KBr method were estimated from the IR spectra by the use of an initial computerized library search and followed by visual inspection of the spectra.
|
Materials and Methods |
|---|
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
) and 236 mixtures. Usually, no more than three components can be detected in one patient sample. The majority of all urinary calculi contain one or more (maximum, three) of the eight most commonly occurring components. Therefore, the 236 mixtures were prepared from these commonly occurring urinary calculus components. The components are as follows: ammonium hydrogen urate (AMUR), brushite (BRUS), carbonate apatite (CARB), cystine (CYST), struvite (STRU), uric acid (URIC), weddellite (WEDD), and whewellite (WHEW; Table 1
). The mixtures were restricted to binary and ternary mixture designs. These so-called constrained and balanced mixtures were prepared in linear ranges of 0–100% with step sizes of 10%. A more detailed description of the preparation of these mixtures can be found elsewhere (13). AMUR was synthesized according to the following brief instructions: 1.68 g of uric acid was suspended in 500 mL water of 37 °C. A 30 g/L ammonia solution was added drop by drop (2 drops/s) until the solution was completely clear. After the solution cooled, the water layer was aspirated. The crystals were subsequently rinsed with water and diethyl ether and dried at 100 °C. With this synthesis, 
0.84 g of AMUR can be obtained, which remains stable for 6 months.
|
CARB and WEDD were obtained from patient samples by a selection based on purity. Purity was established by comparison of IR spectra with those in the Hesse atlas (10) and by standard wet chemical analysis. All mixtures were carefully mixed using a pestle and mortar.
Samples for training and validation of the neural network of the GGN method.
For the development of the neural network, the previously mentioned 236 library mixtures and pure samples of the 8 commonly occurring components were used. Two additional mixtures were added to the set, giving a total number of 248 mixtures. The two extra mixtures were added to make the validation set more representative.
Patient samples for comparison of the KBr and GGN methods.
One hundred consecutively collected urinary calculus samples from 70 males (median age, 56.5 years; range, 5–75 years) and 30 females (median age, 49 years; range, 21–74 years) served for testing the predictive performance of the new GGN method. The majority of them (>95%) were derived from patients treated with extracorporeal shock wave lithotripsy. Before analysis, each whole patient sample was carefully ground using a pestle and mortar. The quantitative composition of each sample was also obtained from the KBr method, using a computerized library search followed by visual inspection of the spectrum. The samples were considered a representative selection of the urinary calculi in our routine practice. The composition of the calculi will be described in greater detail in the Results section.
analytical methods
Standardization.
Before each series, we validated the FT-IR instrument by measurement of a polystyrene transmission standard. Validation comprised wave number positions and absorbances of known IR bands. The linearity of IR analyses of the KBr and Golden Gate assays was tested using a dilution series of URIC at concentrations of 0–100% with step sizes of 10%. URIC was diluted with WHEW, and the area of the URIC band at 1120 cm-1 served for establishment of test linearity. The "runs-test" was used for establishment of significant deviations from a straight line (15). P 
0.05 was considered statistically different.
KBr method.
Pulverized urinary calculus (1.5 mg) was mixed with 180 mg of KBr with a pestle and mortar. From this mixture, 100 mg was taken for the preparation of a urinary calculus KBr tablet at 10 tons of pressure under vacuum for 2 min. A more extensive description can be found in the Hesse atlas (10). The spectra were scanned in the mid-IR region from 4000–400 cm-1 at 4 cm-1 wave number intervals in a Bio-Rad FTS 135 spectrometer equipped with a cooled DTGS detector and Win-IR (Ver. 3.04) software (both from Bio-Rad Laboratories Inc., Spectroscopic Division). A 100-mg KBr tablet was used as a blank for background subtraction. Samples producing weak spectra (absolute difference between absorbance maximum and minimum less than A = 0.25) were reanalyzed using tablets with higher sample:KBr ratios.
The quantitative composition of each sample was estimated by comparison of the recorded spectrum with KBr reference spectra that were stored in a computer library (LITHOS; Bio-Rad). This library contains data of pure components of urinary calculi and 227 mixtures. Its content was similar to a LITHOS library that is used for x-ray diffraction. Win-IR search (Ver. 1.03; Bio-Rad Laboratories Inc., Sadtler Division) served as search engine. This search engine applies the Euclidean distance-matching algorithm to the fingerprint area (2000–400 cm-1) of the absorbance spectra to obtain a spectral hit list. Additional evaluation and interpolation led to an estimate of the quantitative composition of a sample because the first hit is not necessarily the correct one and because even large libraries cannot contain full detail. After the library search, the final composition was obtained by visual inspection of spectral band intensities by two experienced technicians blinded to the results of the Golden Gate method.
General outline of the GGN method.
The so-called Golden Gate is a sample-holding device equipped with a Single Reflection Diamond ATR crystal (Graseby Specac) for measurement of micro samples. The standard ZnSe lens was replaced with a KSR5 lens to enable measurements between 600 and 250 cm-1. Carefully pulverized material (1–2 mg) was applied to the flat surface of the diamond crystal and pressurized at 3 x 108 Pa. The reproducibility of this pressure was guaranteed by use of the built-in pressure restraint of the pressure applicator of the Golden Gate device. The active sampling area of the crystal was 1.13 mm2 (diameter, 0.6 mm). The uniformity of the crystal spreading on the sensing area was controlled by viewing through the looking glass of the pressure applicator of the Golden Gate device. The samples were always measured at room temperature. A Bio-Rad FTS 135 spectrometer, equipped with a cooled DTGS detector and Win-IR software, was used for scanning in the mid-IR region from 4000–400 cm-1 at 4 cm-1 wave number intervals. An empty crystal served for background measurement and blank subtraction. The background spectra were always collected before a series of 10 sample spectra. All training, validation, and test samples were measured in a more or less random order over 6 months. Each spectrum was acquired by coaddition and averaging of 16 interferograms. The Golden Gate crystal was cleaned with water and 960 mL/L alcohol after each measurement.
The NEURANET program (Ver. 3.0; Bio-Rad Laboratories Inc., Spectroscopic Division) was used for quantification of urinary calculi, whose compositions are expressed as mass percentages. The selection of this program was based on earlier studies (13)(14). This program contains two supplementary quantification methods and was particularly developed for interpretation of IR spectra of urinary calculi in the range of 4000–400 cm-1. The first method is a computerized library search, and the second is based on ANN prediction. The library search can be used for quantification of any composition of a calculus, assuming that the components are available at the library. The neural network may be used for more accurate predictions of the composition of urinary calculi, but it is restricted to processing the outcomes of a maximum of 10 components simultaneously. Therefore, the neural network can be used only for quantification of the most commonly occurring components of urinary calculi. For calculi composed of these commonly occurring components, both quantification methods should provide almost the same outcome. In this case, the library search serves as a verification method of the ANN because it can depict the unknown spectrum graphically together with several of the library spectra (stacked or overlaid). For rarely occurring components, the results obtained with the library search must be used. The availability of both quantification methods facilitates the interpretation process. Additionally, the program offers the possibility of adding some simple expert rules to the network-predicted results. These rules may be added to solve problems caused by small but systematic inaccuracies in the network outcome of patient samples. Furthermore, they are used to give an indication that the results from the library search should be used in case of rare components unavailable to the network model and to round the network outcome of each component to the nearest 5%. The expert rules provide a generalization of the quantitative results of future patient samples. The rules may be defined as simple "Basic", such as "IF ... THEN ... ELSE ..." statements.
The neural network engine of the NEURANET program is based on a back-propagation neural network (16). This program contains a three-layer network consisting of an input layer with a number of nodes (neurons) equal to the number of input variables (absorbances at different wave numbers), a single hidden layer with a variable number of nodes, and an output layer with a number of nodes equal to the number of components (maximum, 10). The input nodes are connected to the output nodes via the nodes of the hidden layer. All nodes of the hidden layer have every possible connection with the input and the output nodes. Each connection carries the signal and an individual weight. The final weights in all connections reflect the knowledge of the underlying spectral patterns. The complex of weights may be interpreted as the regression coefficient in a regression analysis. The multiple inputs of a spectrum are converted to a single concentration of a single component. The final set of weights is found by back-propagation. With this method, the neural network is provided with a set of training spectra (samples) with known concentrations and iterates around a loop in which it predicts for each sample the analyte concentrations and compares these to the known concentrations (forward step). Depending on the differences between the calculated and the known outcome concentrations, the weight values will be readjusted (backward step). This happens for each sample of the training set in turn and is repeated many times over the complete data set. The number of iterations (epochs) usually is very large. Before training, the starting weights are randomized between -0.1 and 0.1.
The performance of the network is monitored by looking at the root mean square error (RMSE). The RMSE is calculated by first taking the sum of squared differences between the desired and obtained output values of the training set. The square root is then taken from the average of the sum of squares, which are averaged by the number of outcomes (maximum of 10) and the number of training samples. NEURANET contains several parameters for data preprocessing (e.g., selection of wave number ranges and scaling) and network design (topology). The neural network parameters must be tuned by means of an independent validation set. This validation set is a representative set of spectra from samples of known composition and is used for testing the performance of the network but not for training. The training behavior of the network is monitored by looking at a graph depicting the decrease in and convergences of the RMSE of both the training and the validation sets against the number of epochs. In worse cases, both RMSE curves will diverge instead of converge. To prevent overfitting of the neural network model, the training process is stopped when the RMSE of the validation set is at its minimum value (early stopping rule). In addition to training and validation (used for tuning), the performance of the network should be tested with a test set. More information about neural network processing can be found elsewhere (17)(18).
The NEURANET program enables building of one or more named methods, based on spectroscopic absorbance data. Each method contains a combination of standard information (e.g., description of the components), a spectral library, a trained neural network model, and some expert rules. After selection of a method and of a spectrum from the file list, the program automatically performs a library search, network predictions, and expert rule filtering for prediction of the outcome of a sample with an unknown composition. The whole combination of the analysis with the Golden Gate sampling device and the final NEURANET model is called the GGN method.
development of the library and neural network of the ggn method
The 261 samples, composed of 25 pure components and 236 mixtures, were analyzed with the Golden Gate sampling device, and their spectra were added to the library of the NEURANET program. They were stored at 16 cm-1 resolution intervals of the 4000–400 cm-1 analysis range. A spectral range was defined for searching in the 3700–450 cm-1 range with the correlation-matching algorithm. The spectra of 248 pure components and mixtures of AMUR, BRUS, CARB, CYST, STRU, URIC, WEDD, and WHEW were recorded with the Golden Gate device. Of these, 199 were used as a training set for the neural network. This training set served for the construction of a network model that has a mapping (topology) suitable for the analysis these eight components (eight output neurons) in the unknown samples. The remaining 49 spectra were used as a validation set. A spectral range from 1840 to 448 cm-1 with 16 cm-1 resolution intervals was selected for neural network processing, giving rise to 88 input neurons. The network topology parameters for training the neural network were optimized by monitoring the RMSE of the validation set. The RMSE values for both the training and the validation sets were graphically depicted to check potential overfitting of the neural network model. A more extensive description of network processing in relation to urinary calculus analysis can be found elsewhere (14). A few expert rules were added to the network-predicted data for further optimization of the results of unknown samples. These rules were added as a result of small but structural differences in composition found between visual inspection of the Golden Gate spectra by two technicians and the network-predicted results of the patient samples (test set). The rules were added without any foreknowledge of the results from the KBr method.
data processing and statistics
The results of the KBr and GGN methods were compared using Bland–Altman agreement plots (19) for the eight commonly occurring components and a combination of these. With these agreement plots, the individual differences (AMUR%[KBr]n - AMUR%[GGN]n, CYST%[KBr]n - CYST%[GGN]n of sample n, and so forth) of both methods are calculated and plotted against the individual mean results (e.g., mean of AMUR%[KBr]n and AMUR%[GGN]n of sample n) of both methods. The bias (mean of the individual differences between the GGN and KBr methods) and the 95% agreement limits (1.96 SD of the differences between both methods) are summarized in Table 3
. The bias and 95% agreement level were calculated for the eight components separately and for a combination of them by taking the individual calculated differences of the eight components together. The bias and 95% agreement limits of the combination of the eight components are shown in an agreement plot in Fig. 3
.
|
|
|
Results |
|---|
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
. Fig. 1B
shows a transmission spectrum of the same sample obtained with the Golden Gate device. Analysis of the URIC dilution series with the runs-test showed that the IR intensities of both the KBr and GGN assays were linear at 0–100%.
|
composition of urinary calculi with the KBr method
Two of the 100 analyzed patient samples produced weak spectra and were removed from the data set because of insufficient sample material for reanalysis. The quantitative composition of the remaining 98 patient samples, as analyzed with the KBr method, revealed that 92 of them contained at least one of the eight commonly occurring components. All 92 patient samples were single components or binary or ternary mixtures of one of these eight components. The majority of them contained calcium oxalate (WHEW and WEDD) and/or CARB. The detection frequency of each of the eight components in the 92 samples was as follows: AMUR, 1.1%; CYST, 1.1%; URIC, 3.3%; STRU, 4.3%; BRUS, 13%; CARB, 48%; WEDD, 70.7%; and WHEW, 75%. The percentages of urinary calculi that contained one, two, or three of these components were 14%, 54%, and 32%, respectively. Six of the 98 urinary calculi contained less frequently occurring components. One consisted of quartz, two of uric acid dihydrate, and three of a fatty substance. Two of the latter were highly similar to feces, whereas the other was similar to palmitic acid.
development of the ggn method
Neural network.
After repeated, batch-automated training of the neural network with different topologies, a final topology was found. Each training session took 20 min for each topology. The final topology had eight hidden neurons. With this topology, the RMSE steadily decreased to a minimum value of 2.3% for the validation set (Fig. 2
). This minimum value was reached after a training of 54 000 epochs (cycles). At this number of epochs, the error of the training set was 1.5%.
|
Intermediate analysis of composition of urinary calculi.
The 98 patient samples, which were also analyzed by the KBr method, were analyzed with the Golden Gate device. The composition of each calculus was estimated with library searches and neural network prediction of the intermediate GGN method. Computerized estimation of the composition of a single sample with the GGN method was obtained within 1 s by means of a simultaneous library search and network prediction followed by expert-rule filtering. The composition of 6 of the 98 patient samples could not be estimated by the ANN because these samples did not contain any of the eight commonly occurring components available in the network model. Four of these samples could be detected with library searches of the GGN method in the first hit (searched from 3700 to 450 cm-1). As with the KBr method, one was found to contain quartz, two contained feces, and one contained a component similar to palmitic acid. The composition of the remaining two, which contained uric acid dihydrate according to the KBr method, could not be resolved by library searches with the GGN method because the spectrum of uric acid dihydrate was not available in this library.
Addition of expert rules.
As a result of visual inspection of the Golden Gate spectra and the outcome of the neural network predictions of the 92 patient samples (test set), 14 simple expert rules were added to the GGN method (Table 2
). After addition of all expert rules to the GGN method, the composition of each of the 92 patient samples, as analyzed with the Golden Gate device, was reestimated with the final GGN method. The results of the final GGN method were used for comparison with those of the KBr method.
|
comparison of the KBr and the ggn methods
The agreement between the KBr and GGN methods, as obtained from the Bland–Altman plots of the results of 92 patient samples, is shown in Table 3
. The Bland–Altman plot of all patient results is shown in Fig. 3
. The plot compares the results for all eight components obtained from the 92 patient samples, analyzed with the KBr and the GGN methods. The dashed lines express the 95% confidence interval of the differences between both methods. Of the 92 samples, 2 consisting of WHEW + WEDD and 1 consisting of CARB + STRU showed 20% difference between both methods (see Fig. 3
). Because each of the three samples contained mixed stones composed of two concomitantly occurring components, an increased amount (percentage) of one component relative to the other sample produced an equal decrease of the amount of the other component. Therefore, the Bland–Altman plot shows six data points at 20% difference between the methods. For example, the sample containing CARB and STRU was composed of 60% CARB and 40% STRU measured with the KBr method, whereas it contained 80% CARB and 20% STRU measured with the Golden Gate method. The resulting data points (x,y) of the Bland–Altman plot are (70,20) and (30,20).
|
Discussion |
|---|
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
Visual inspection of the IR spectra made clear that KBr spectra (Fig. 1A
) have more definite bands than those recorded with the Golden Gate ATR device (Fig. 1B
). The Golden Gate ATR device also yields spectra with less absorbance intensities at higher wave numbers when compared with the traditional transmission spectra (Fig. 1A
). The underlying cause is different sample radiation penetration depths at different wave numbers. It did not influence the interpretation of the spectra obtained with the Golden Gate assay, probably because of the predominant use of the bands with sufficient spectral definition in the 2000–400 cm-1 region (fingerprint area).
development of the neural network model
From a previous study (14), we found that the compositions of urinary calculi using neural network prediction were similar to the results obtained with PLS regression. In both cases, the urinary calculi were analyzed with IR spectroscopy using KBr tablets. The PLS regression and the neural network model were developed for quantification of mixtures of WHEW, WEDD, and CARB, whose incidence rate in urinary stones is 80% in Western countries (20). The RMSE values of the validation sets were 1.7% for PLS regression and 1.6% for network prediction, respectively. The previous study also described the development of a neural network model trained with the eight most commonly occurring components. This neural network model has been used successfully for several years in our laboratory (UHG). On the basis of these findings, we concluded that for quantification of the composition of urinary calculi, neural networks would be as useful as or better than linear models such as PLS regression. We therefore developed a new neural network model with the NEURANET program after replacing the KBr sampling device by the Golden Gate accessory.
Training of the ANN went remarkably well, despite the previously mentioned relatively poor spectral band definition in the fingerprint area compared with the KBr method. Although the NEURANET program contains several facilities to make network training rather simple, there are several difficulties in applying ANN models. The ANN parameters (topology) often are difficult to estimate, and large training sets often are needed. The number of training samples should be at least more than the number of input units (88 in our case), but the required number of training samples also strongly depends on the noise in the targets and the complexity of the adaptation of the network to the target function. The absorbance values of IR spectra are linearly related to the concentration (Beer–Lambert law). Therefore, a relatively small number of learning samples (n = 199) was needed for training. We gave special attention to the risk of overtraining (overfitting). In this case, the network loses generalization (robustness) and will adapt (learn) to unimportant spectral features, such as noise. Overfitting (and underfitting) was monitored by looking at the RMSE errors of the validation and training sets (Fig. 2
). If the validation error became much higher than the training error, the network was probably overfitted and another topology was applied. In addition, the robustness of the neural network model is important and can be tested by retraining the neural network with different initial weights each training session, providing that the other conditions are left unchanged (e.g., topology settings, training, and validation sets). If the RMSE values at a certain number of epochs show large differences for the different training sessions, the network model can be considered unstable. The final criterion for assessing the network model was the comparison of the results from the independent test set with the results obtained with the reference method.
Several heuristics exist for the choice of the starting values of many of the topology parameters. However, because the optimal parameter settings strongly depend on the nature of the problem and on the chosen representation of the input and output objects, it is not safe to rely exclusively on heuristics. Therefore, an operator must have sufficient knowledge of network training to select the topology parameters (e.g., number of hidden neurons) and must interpret the numerical and graphical network outcome. With NEURANET, consecutive unattended training with different topologies was possible when several topologies were set out in advance. After training, estimation of the composition of patient samples is very fast when the stored ANN method is used.
More detailed information about the theoretical background of ANNs is out of the scope of this report, but can be found elsewhere (17)(18).
development of the expert rules
Urinary calculi are always composed of pure components or binary or ternary mixtures. For those components absent in the sample (e.g., five absent in case of a three-component calculus), small positive or negative numbers may occur in the network outcome of patient samples on a regular base (Table 4
). This is a consequence of network training, which will always predict the outcome of the eight components simultaneously, whether or not they are present in the sample. Because the total outcome of any sample is always 100% (Table 4
), a small outcome (percentage) may occur for those components absent in the sample. These inaccuracies may occur because no prediction is perfect, the samples may contain trace amounts of impurities caused by their passage through the urinary tract, and some samples produce rather noisy spectra. In this last case, the network assumes detection of small amounts of a component characterized by a great number of spectral bands (e.g., AMUR). Therefore, a few expert rules were defined (see Table 2
). In essence, the expert rules can be considered as automated corrections of the network outcome, which otherwise would have been made manually by expert technicians after visual inspection of the IR spectra. Some of these rules are counterparts of each other, describing almost the same type of correction (e.g., amurcheck1 and amurcheck2). Another rule, named whewwedd3, seems to be rather complex, but only assigns the smallest oxalate outcome (Whew or Wedd) to the largest one, providing that both oxalate outcomes are positive and <3.5%. The rationale is not a physical/chemical one, but only a small correction. If this rule was not applied, both oxalate outcomes would be forced to zero by rounding and normalization, in spite of the fact that a small amount of oxalate is present in the sample.
|
Except for normalization of the network outcome to the nearest 5%, no special expert rule was applied to patient sample A (Table 4
). Sample B shows somewhat inaccurate results of both calcium oxalates (WHEW and WEDD). WHEW turned out to be predominant relative to WEDD. Because this has happened several times, an expert rule was defined by simply adding the values of both calcium oxalates and forcing the value of WEDD to zero. This rule was defined as follows:
 |
This rule yields 0% WEDD and 4% WHEW. After normalization of sample B, the final composition of the expert system was 5% WHEW and 95% CARB (Table 4
). This outcome reflected the real composition of this sample, which was based on careful visual inspection of the band intensities of the KBr spectrum by a trained technician.
method comparison
The 92 consecutive samples used for method comparison were regarded as a representative selection of urinary calculi in our daily practice. They had similar frequency distribution of components and number of components per sample, compared with historical data (not shown). X-ray diffraction is occasionally recommended as a reference method for urinary calculus analyses. However, x-ray diffraction cannot adequately detect amorphous substances (3). CARB is, for example, sometimes overlooked, but can be detected by a simple CO2 test following acidification with HCl. Quantitative analysis of CARB may, however, be difficult. We therefore decided to compare the GGN results in both an analytical and a managerial sense with those obtained by an IR method with KBr tablets. This method was routinely used at the time of the study and was to be substituted with a less time-consuming and more robust analytical method.
The bias of the outcome of the KBr and GGN methods of the 92 patient was significantly different from zero for BRUS, WHEW, and WEDD (Table 3
). The small bias of BRUS (-0.8%) seems irrelevant. The biases for WHEW (-2.8%) and WEDD (3.8%) are small and carry different signs, probably related to their concomitant occurrence in urinary calculi. The 95% levels of agreement of WHEW, WEDD, and CARB were >10% (Table 3
). These components often occur concomitantly in a single sample, causing complex spectral patterns. The 95% level of agreement of all results was 9%. This value should be taken as an indication because it is statistically not correct to base such calculations on mutually dependent variables (each sample occurs 8 times). Only 3 of 92 patient samples exhibited maximum differences of 20%. These differences occurred consistently in samples that contained two rather similar components (WHEW + WEDD and CARB + STRU). It is not known what analytical precision and bias are relevant in terms of the prevention of urinary calculus recurrence. We nevertheless consider the encountered differences minor and possibly irrelevant with respect to the ultimate (dietary) advice.
Apart from adequate quantification of the eight commonly occurring components, the library search in the GGN method enabled detection and quantification of rarely occurring components in four samples. This feature may be further developed by the addition of other components to the library, such as uric acid dihydrate, in the near future. On the other hand, a library search may be used for verification of network results. However, it may sometimes be somewhat difficult to establish an accurate quantitative composition of a sample in this way. This is illustrated by the results from the first and second hits obtained with a library search of patient A in Table 4
.
In conclusion, the GGN seems superior to the KBr assay because of its smaller sample size, because there is no need for sample pretreatment except for grinding, the turnaround time is shorter, and no time is lost because of KBr tablet breakage (Table 5
). The GGN method does, however, require higher initial investment because of the Golden Gate ATR sampling device. Because no sample pretreatment is needed, different brands of FT-IR spectrometers give similar spectra under equal local conditions (e.g., temperature and sample pressure), and the chemical composition of urinary calculi is similar in most developed countries, it would be interesting to investigate whether the GGN method could be transferred to other laboratories without retraining the neural network with local data. This, however, awaits confirmation. The required expert knowledge for spectral interpretation is minimized by the use of the ANN and library, but visual inspection remains necessary.
|
|
Acknowledgments |
|---|
|
Footnotes |
|---|
|
References |
|---|
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
