Smelling Renal Dysfunction Via Electronic Nose

Annals of Biomedical Engineering, Vol. 33, No. 5, May 2005 (© 2005) pp. 656–660 DOI: 10.1007/s10439-005-1438-2

Smelling Renal Dysfunction via Electronic Nose ANDREAS VOSS,1 VICO BAIER,1 RENATE REISCH,2 KATHARINA VON RODA,3 ¨ PETER ELSNER,4 HORST AHLERS,2 and GUNTER STEIN3 1

Department of Medical Engineering, University of Applied Sciences Jena, Germany; 2 Jenasensoric e.V., Am Planetarium 5, Germany; 3 Department of Internal Medicine III, Division of Nephrology, Friedrich-Schiller-University of Jena, Germany; and 4 Department of Dermatology and Allergology, Friedrich-Schiller-University of Jena, Germany (Received 14 May 2004; accepted 19 November 2004)

to be low, whereas urea concentration in the sweat fluid was found to be present at a much higher concentration than the serum level.2 Patients do not smell their uremic odor themselves because their ability to smell is severely impaired and related to the degree of renal impairment and the degree of accumulation of uremic toxins.4 Olfactory evaluation can provide diagnostic clues, guide the laboratory evaluation, and help in the choice of immediate and appropriate therapy.10 In recent years the progress in sensor systems (electronic noses) performance opened the way to the possibility of fast and simple analysis of odors in many fields, and, recently in medicine for analyzing volatiles secreted outside the human body to get information about the health status.9,13 Our present study shows the application of an electronic nose system based on doped semiconductor metal oxide gas sensors for investigating human body odor in patients with different stages of renal insufficiency and the relation between body odor and medical standard parameters.

Abstract—The human body odor plays an important role in social communication in various situations, like the olfactory identification of partners and relatives as well as in parents–child interactions. In patients with renal dysfunction the compound of sweat and volatile gases is changed because of the limited ability for removing metabolic products from the blood. The regulation of electrolyte composition and acid–base balance are also altered so that the body odor of these patients may be significantly influenced by these disorders. We show the ability of an electronic nose to detect changes in the human body odor in consequence of renal dysfunction by reducing multivariate sensor signals with principal component analysis to its first and second principal odor component (POC). All healthy subjects could clearly be distinguished from patients with renal failure using quadratic discriminant analysis, whereas a correct classification of 95.2% (98.4% using 1st– 3rd POC) of patients between end stage renal failure and chronic renal failure was found. This methodology of analyzing human body odor may also provide new approaches for investigating symptoms of renal failure and for diagnosing other diseases of internal or cutaneous origin.

Keywords—Body odor, Electronic nose, Principal component analysis, Discriminant analysis.

METHODS INTRODUCTION

Patients

More than a hundred years ago “malodorous perspiration” in humans with malfunctioning general condition was first reported.7 Also human kidney diseases which are accompanied by a significantly reduced glomerular filtration rate influence breath as well as body odor. Simenhoff et al.11 also showed that uremic breath reflects the systemic accumulation of potentially toxic volatile metabolites. These metabolites change the composition of secretion products from exocrine glands and therefore the compounds of volatile gases released by the skin, which characterizes the individual body odor. The concentration of creatinine and uric acid in the sweat fluid was shown

The body odor was examined in 42 patients with end stage renal failure (dialysis patients—DP, mean age 59 ± 15 years) who underwent regular hemodialysis three times a week, 20 patients with chronic renal failure (CRF, stages 3–5 in accordance with the practice guidelines of the Kidney Disease Outcomes Quality Initiative—K/DOQI, mean age 63 ± 15 years) and 11 healthy controls (CON, mean age 47 ± 14 years). Blood samples were collected for determination of biochemical parameters immediately before investigating the body odor. Table 1 shows the mean values and standard deviation of the biochemical parameters. Measurements were performed between 7am and 10am in CRF and CON, in DP in the first 30 min of dialysis session. The sensor head (volume 5 ml) of the electronic nose8 was placed on patients leg as shown in Fig. 1. Three sensor signals were recorded and corrected/calibrated to

Address correspondence to Prof. Dr. A. Voss, University of Applied Sciences Jena, Carl-Zeiss-Promenade 2, 07745 Jena, Germany. Electronicmail: [email protected]

656 C 2005 Biomedical Engineering Society 0090-6964/05/0500-0656/1


Smelling Renal Dysfunction via Electronic Nose

657

TABLE 1. Biochemical parameters. DP Mean Creatinine (µmol/l) Potassium (mmol/l) Phosphorus (mmol/l) pH Albumin (g/l) Haemoglobin (mmol/l) Haematocrit Serum iron (µmol/l) iPTH (ng/l) 25-OHD (ng/ml) 1,25-(OH)2-D3 (pg/ml) Blood sugar (mmol/l) Calcium∗ phosphate product (mmol/l2 ) Creatinine clearance (ml/min∗ 1.73 m2 )

887 5.17 2.06 7.40 59.5 6.79 0.32 12.1 251.3 16.7 8.2 5.80 4.53

CRF SD 258 0.77 0.49 0.04 6.0 0.82 0.04 3.6 238.0 8.1 8.2 2.09 1.16

Mean 403 4.72 1.59 7.31 35.7 7.01 0.35 12.9 225.9 24.7 20.9 8.15 3.71 0.33

CON SD

123 0.49 0.37 0.06 4.9 0.90 0.04 3.1 136.7 13.0 9.5 3.30 1.03 0.18

Mean 80 4.03 1.18 7.30 41.4 8.97 0.43 23.8 26.1 29.6 64.1 5.72 2.81

SD

p

15 0.20 0.20 0.05 4.9 0.97 0.05 8.8 12.7 14.0 16.9 2.02 0.52

∗ † ‡ † ‡ ∗ † ‡ ∗ † ∗ † ‡ † ‡ † ‡ † ‡ † ‡ † ∗ † ‡ ∗ †

Note. Mann–Whitney Test, significant parameters (p < 0.01) are indicated as follows: between DP and CRF (∗ ), DP and CON († ), CRF and CON (‡ ).

compensate for the reading from ambient air before principal component analysis has been carried out. Electronic Nose The measuring system8 consists of an array of three thick-film metal oxide based gas sensors with heater elements. Each of the sensors has a slightly different sensitivity to various odorant molecular types. Molecular components in the gas phase are allowed to pass by the sensors and the interactions between molecules and the sensors involve reactions with oxygen on the sensor surface, which lead to a change of the free charge carrier concentrations in the conducting metal oxide. Advantages of these sensors are high sensitivities and long duration. The integrated platinum heater allows the sensor to operate from 200◦ C up to

385◦ C to adapt the sensor’s sensitivity. The sensor layers are sensitive to oxidizing and reducing volatile gases such as methane, butane, alcohols, ketones, carbon monoxide, nitrogen oxide, ammonia and humidity, which cause the sensor resistance to change. The sensor resistance was measured in the range from 200◦ C to 385◦ C, which lasted approximately 2 min. Ten measuring cycles were recorded per patient to yield an overall measurement time of 20 min. For the assessment of the body odor only the last measuring cycle was used, since after 18–20 min the gas milieu in the sensor was the best representation of the substances released by the skin. Each measurement was accompanied by a calibration procedure in which a calibration factor cf from the ambient air was determined. Therefore, the three corrected sensor signals (CSS) were calculated from the sensor resistance (R1−3 ) and gain control resistance (Rk ) in the following: CSS1 =

R3 R1 − cf 1 , Rk Rk

CSS3 =

R2 R3 − cf 3 Rk Rk

CSS2 =

R1 R2 − cf 2 , Rk Rk

Multivariate Analysis

FIGURE 1. Measurement setup of the electronic nose system.

To discriminate between complex body odors in the different groups the corrected sensor signals were analyzed using Principal Component Analysis6 (PCA). This is an effective method to reduce our multidimensional data space (3 sensors, each operating at 185 different temperatures) to its main components and therefore improves the human perception ability of the data. PCA determines the linear combinations of the sensor values such that the maximum variance between all data points can be obtained in mutually orthogonal

658

VOSS et al.

FIGURE 2. First and second principal odor components.

dimensions. This results in the largest variance between sensor values from the first principal odor component and produces decreasing magnitudes of variance from the second to the third odor components and so on. To associate these odor components with clinical parameters Pearson’s correlation coefficients were calculated. To assess dissimilarities in the body odor we reduced the complete sensor signals to its first and second principal odor component. Quadratic discriminant analysis3 was then applied as statistical analytical technique to test the ability of the electronic nose for separating the odor of healthy subjects from those of patients with renal dysfunction. The variables used for this purpose are the first and second principal odor components. RESULTS The 1st and 2nd principal odor components (POC) for each subject are drawn in Fig. 2. They account for 91% of the variance in the data and 1st–3rd POC even for 97% variance because of the expected high correlation between sensor values at different temperatures. The cluster of healthy controls (oval) can be separated from all renal patients by quadratic discriminant analysis to 100%. It is obvious that POCs also differ between CRF (square points) and DP (triangle points)—the group with the highest intra-group

variance—but show a small overlapping area. Quadratic discriminant analysis of the 1st and 2nd POC from the two groups of renal patients reveals a correct classification of 95.2% (three patients are misclassified), whereas correct classification increases to 98.4% (only one patient is misclassified) using 1st–3rd POC. Table 2 shows all significant correlation coefficients between biochemical parameters and 1st and 2nd POCs. The left column of this table represents the correlations including all three groups and the right column the results where only the two groups of renal patients are considered. The reason for this is to account for (1) the different odor compounds between healthy subjects and all renal patients, which mainly caused the variance in the 1st POC, and (2) the different odor compounds in CRF and DP that mainly cause the variance in the 2nd POC. As can be seen from the left column in Table 2 (considering CRF, DP and CON) the 1st POC correlates significantly with 1,25-(OH)2-D3 (r = 0.58), serum iron (0.57), creatinine (r = −0.52) and other clinical parameters. This can easily be attributed to the significant differences between renal patients and CON as shown in Table 1 and the already mentioned variance in the 1st POC caused by the different odor compounds in renal patients and CON. Therefore, the correlations with the 1st POC disappear (right part of Table 2) if only the two groups of renal patients are considered for correlation analysis.

Smelling Renal Dysfunction via Electronic Nose

659

TABLE 2. Significant correlation coefficients (p < 0.01) between the first two principal odor components and clinical parameters.

Creatinine (µmol/l) Potassium (mmol/l) Phosphorus (mmol/l) pH Albumin (g/l) Haemoglobin (mmol/l) Haematocrit Serum iron (µmol/l) iPTH (ng/l) 25-OHD (ng/ml) 1,25-(OH)2-D3 (pg/ml) Blood sugar (mmol/l) Calcium∗ phosphate product (mmol/l2 )

Correlations, including all three groups

Correlations, only renal patients considered

1st POC

2nd POC

1st POC

−0.52 −0.48 −0.43 −0.30 −0.41 0.35 0.38 0.57

−0.42

−0.52

−0.51 −0.65

−0.55 −0.69

2nd POC

0.33

0.58

0.40 0.51

−0.38

Note. No entry means not significantly correlated.

The 2nd POC, whose variance is mainly caused by varying odor compounds of DP and CRF shows the highest significant correlation with albumin (r = −0.69), pH (r = −0.55) and creatinine (r = −0.52). Table 1 shows that these parameters differ significantly between the two groups. DISCUSSION The present results show the ability of an electronic nose system to classify human body odor of patients with different stages of renal insufficiency (Fig. 3). All healthy subjects could be discriminated from renal patients using the first two principal odor components and quadratic discriminant analysis. Furthermore, a correct classification of 95.2% (1st and 2nd POC) and 98.4% (1st–3rd POC) was achieved in discriminating different stages of chronic renal failure, especially DP and CRF. Nevertheless, the aim of this study was not to find a new method for the identification of renal failure patients, we

FIGURE 3. Correct classification rate in each two groups comparison with quadratic discriminant analysis using the 1st and 2nd POC (or 1st–3rd POC, respectively).

rather showed the applicability for investigating renal failure patients’ odor with an electronic nose. This methodology may also provide an approach for identifying symptoms of chronic renal failure. Pruritus is one such symptom that is characterized by an intense itching sensation that produces the urge to rub or scratch the skin to obtain relief and cannot be characterized in the laboratory so far.14 We speculate that the composition of perspiration should be changed in these patients. If principal odor components measured with an electronic nose could help in the classification of renal failure patients with and without pruritus, it should be possible to link this odor components with clinical correlates that may improve the understanding of the pathogenesis of this symptom. Besides the identification of renal dysfunction it may be possible to apply our methodology to the diagnosis of other diseases of internal or cutaneous origin. In our analysis we did not investigate the influence of the measuring site, which was shown to affect the composition of sweat and therefore body odor.12 This, however, might yield further improvements in the discrimination between different stages of renal insufficiency. Concentrations of electrolytes, glucose and urea in sweat were reported to increase with age, whereas concentrations of lactate, total protein and total lipids, however, were not age-dependent.1 Furthermore, Haze et al.5 reported an age-related change of body odor. Even though in our investigation age differed significantly between renal patients and healthy subjects, the 1st and 2nd principal odor components were not correlated with age and therefore age differences did not contribute to the high discrimination rate between the groups. In conclusion, we showed that the application of an electronic nose system for analyzing human body odor allowed the distinction between different stages of renal dysfunction. Now it has to be investigated if this methodology may also be used to identify symptoms of chronic renal failure,

660

VOSS et al.

which could improve the understanding of its underlying pathological processes. ACKNOWLEDGEMENT This work was partly supported by grants from the Federal Ministry of Education, Science, Research and Technology BMBF 0312704C. REFERENCES 1

al-Tamer, Y. Y., and E. A. Hadi. Age dependent reference intervals of glucose, urea, protein, lactate and electrolytes in thermally induced sweat. Eur. J. Clin. Chem. Clin. Biochem. 32(2):71–77, 1994. 2 al-Tamer, Y. Y., E. A. Hadi, and I. I. al-Badrani. Sweat urea, uric acid and creatinine concentrations in uraemic patients. Urol. Res. 25(5):337–340, 1997. 3 Croux, C., and C. Dehon. Robust linear discriminant analysis using S-estimators. Can. J. Statist. 29:473–492, 2001. 4 Griep, M. I., P. Van der Niepen, J. J. Sennesael, T. F. Mets, D. L. Massart, and D. L. Verbeelen. Odor perception in chronic renal disease. Nephrol. Dial Transplant. 12(10):2093–2098, 1997. 5 Haze, S., Y. Gozu, S. Nakamura, Y. Kohno, K. Sawano, H. Ohta, and K. Yamazaki. 2-Nonenal newly found in human body odor

tends to increase with aging. J. Invest. Dermatol. 116(4):520– 524, 2001. 6 Jackson, J. E. A User’s Guide to Principal Components, pp. 1–25, Wiley, 1991. 7 Jaeger, G. Stoffwirkung in Lebewesen. Grundgesetzliches f¨ur Lebenslehre und Lebenspraxis. Leipzig. Ernst G¨unther’s Verlag, 1892. 8 L¨ober, G., and H. Ahlers. Patent DE 101 09 148 A 1. Anordnung zur Detektion von K¨orperfl¨ussigkeiten und—bestandteilen. 9 Mantini, A., C. Di Natale, A. Magagnano, R. Paolesse, A. Finazzi-Agro, and A. D’Amico. Biomedical application of an electronic nose. Crit. Rev. Biomed. Eng. 28(3–4):481–485, 2000. 10 Senol, M., and P. Fireman. Body odor in dermatologic diagnosis. Cutis 63(2):107–111, 1999. 11 Simenhoff, M. L., J. F. Burke, J. J. Saukkonen, A. T. Ordinario, and R. Doty. Biochemical profile or uremic breath. N. Engl. J. Med. 297(3):132–135, 1977. 12 Taylor, R. P., A. A. Polliack, and D. L. Bader. The analysis of metabolites in human sweat: analytical methods and potential application to investigation of pressure ischaemia of soft tissues. Ann. Clin. Biochem. 31(Pt 1):18–24, 1994. 13 Thaler, E. R., D. W. Kennedy, and C. W. Hanson. Medical applications of electronic nose technology: review of current status. Am. J. Rhinol. 15(5):291–295, 2001. 14 Virga, G., S. Mastrosimone, G. Amici, G. Munaretto, F. Gastaldon, and A. Bonadonna. Symptoms in hemodialysis patients and their relationship with biochemical and demographic parameters. Int. J. Artif. Organs 21(12):788–793, 1998.