2884 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 57, NO.

12, DECEMBER 2010

Direct Identification of Bacteria in Blood Culture

Samples Using an Electronic Nose
Marco Trincavelli, Member, IEEE, Silvia Coradeschi, Amy Loutfi*, Bo Söderquist, and Per Thunberg

Abstract—In this paper, we introduce a method for identifica- before a final identification can be made. Thus, finding diag-
tion of bacteria in human blood culture samples using an electronic nostic methods that provide fast detection of the presence of the
nose. The method uses features, which capture the static (steady bacteria and identification of its type, thereby, allowing proper
state) and dynamic (transient) properties of the signal from the
gas sensor array and proposes a means to ensemble results from antibiotic treatment can provide an increased benefit to the pa-
consecutive samples. The underlying mechanism for ensembling is tient as well as help to reduce cost to the health care system.
based on an estimation of posterior probability, which is extracted In the last decade, compact gas sensors have been integrated
from a support vector machine classifier. A large dataset repre- into an instrument called an electronic nose, which can pro-
senting ten different bacteria cultures has been used to validate vide fast (order of minutes) and nonintrusive measurement of
the presented methods. The results detail the performance of the
proposed algorithm and show that through ensembling decisions gaseous agents. Within the medical field, electronic noses have
on consecutive samples, significant reliability in classification ac- been applied to the identification of bacterial agents in different
curacy can be achieved. contexts ranging from upper respiratory diseases [1], urine sam-
Index Terms—Bacteria identification, electronic nose, sepsis. ples [2], and ear, nose, and throat bacteria [3]. The challenge
of developing an electronic noses for bacteria identification in
blood cultures depends on a number of factors. First, the way in
which headspace sampling system introduces the samples to the
I. INTRODUCTION sensor array have been shown in comparative studies to have an
EPSIS, also known as blood poisoning of septicaemia, is effect on identification rates, where samplers that control tem-
S caused by the presence of microorganisms, predominantly
bacteria in circulating blood. Rapid administration of efficient
perature and humidity provide better performance results [4].
Second, the blood in which the bacteria is sampled may differ
antibiotic treatment is crucial as sepsis can result in septic shock, from person to person, and thus, identification should be inde-
multiple organ dysfunction, and even death. The current stan- pendent from the medium itself. Third, the postprocessing of the
dard procedure for diagnosis involves routine microbiological signals should be suited for the characteristics of the signals and
blood cultures. Such procedures can take at least 36 h to sev- application. In applications with a larger number of sensors, it
eral days before diagnosis can be made. Typically, automated is important to extract the relevant features from the signals that
blood-culture-monitoring systems are first used to incubate a can be used for further processing. Although, trials have been
blood culture and monitor the production or reduction of gases made with a single sensor type for identification of bacteria in
(this is normally done via nonintrusive methods for detection of blood cultures [5], today’s off-the-shelf electronic nose devices
CO2 ). Once a sample indicates a change in gas tension, a lab contain anywhere from 4 to 32 sensors in an array. This allows
technician will culture the sample on plates for further identi- for a wider range of odors to be detected.
fication. This secondary subculturing may require up to 36 h In this paper, a new approach for classifying bacteria with
an electronic nose is presented. The method evaluates the suit-
ability of a given sample for classification by representing the
output from a support vector machine (SVM) with a posterior
Manuscript received December 13, 2009; revised March 2, 2010; accepted probability estimation. This estimation is ensembled across ten
March 27, 2010. Date of publication May 10, 2010; date of current version
November 17, 2010. This work was supported by European Union structural consecutive responses of the same sample in order to make the
funds from NovaMedTech. Asterisk indicates corresponding author. classification more reliable. An electronic nose containing 22
M. Trincavelli and S. Coradeschi are with the Center for Applied Au- gas sensors with partial and overlapping selectivities and an au-
tonomous Sensor Systems, School of Science and Technology, Örebro
University, Örebro, SE-70182, Sweden (e-mail: [email protected]; tomatic headspace sampler is used to regulate the samples. The
[email protected]). data processing methods presented consist of extracting features
*A. Loutfi is with the Center for Applied Autonomous Sensor Systems, that reside on the static (steady state) and dynamic (transient)
School of Science and Technology, Örebro University, Örebro, SE-70182,
Sweden (e-mail: [email protected]). properties of the signal. These features are fed into a SVM and
B. Söderquist is with the Örebro University Hospital, Örebro, to the ensembling algorithm. The mechanism of ensembling is
Sweden, and also with the Örebro University, Örebro, SE-70182, based on treating the posterior probabilities as a random sample
Sweden (e-mail: [email protected]).
P. Thunberg is with the Department of Medical Physics, Örebro Univer- and estimating the 95% confidence interval for the mean of the
sity Hospital, Örebro, Sweden, and also with the Örebro University, Örebro, posterior of each class. A mean with significant superior confi-
SE-70182, Sweden (e-mail: [email protected]). dence interval for a class is disjoint and above all the others then
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. classification is performed (assigning the sequence of samples
Digital Object Identifier 10.1109/TBME.2010.2049492 to that class); otherwise a rejection is declared. Identifications

0018-9294/$26.00 © 2010 IEEE

TRINCAVELLI et al.: DIRECT IDENTIFICATION OF BACTERIA IN BLOOD CULTURE SAMPLES USING ELECTRONIC NOSE 2885

TABLE I The sampling cycle, as in most e-nose-based systems, is com-

BACTERIA SPECIES CODE AND TAXONOMY
posed of three phases: baseline acquisition, odor sampling, and
recovery to initial state. In the baseline acquisition phase, the
sensor array is exposed to a reference gas (air in this case) for
10 s, and the value of the sensors is recorded. During the odor
sampling, phases the headspace in the analysis bottle is injected
into the sensor chamber for 30 s. After this, the sensors are ex-
posed again to the reference gas for 260 s in order to recover
the value the sensors had during the baseline acquisition phase.
The total length of the sampling cycle is 5 min.
The sampling cycle is repeated ten times in a row, and we refer
to a series of ten consecutive sampling cycles as a measurement.
The whole dataset (i.e., 1200 sample cycles) is composed of 12
measurement sequences for each of the ten bacteria types: six
measurement sequences with a first batch of bacteria cultures
and six more measurement sequences with a second batch, one
are made among ten typical bacteria cultures leading to sepsis
week later. Blood samples within a batch came from the same
by directly sampling after the first-stage incubation indicates a
source, and different sources were used between batches.
positive culture growth.

III. PATTERN RECOGNITION ALGORITHM

II. EXPERIMENTAL METHOD
The pattern recognition algorithm is articulated into five
A. Sample Preparation phases, namely feature extraction, dimensionality reduction,
Bacteria isolates obtained from clinical samples were used. classification, posterior probability estimation, and ensembling
The bacteria were subcultured on blood agar plates and a bac- decisions. The following sections describe each of the phases of
terial suspension solution was adjusted at turbidity of standard the algorithm.
MacFarland 1.0 (3 × 108 cfu/mL) in NaCl Solution. After that
2 ml of the bacteria suspension were added to an aerobic blood A. Feature Extraction
culture bottle (BD BACTEC Plus + Aerobic/F) together with In several studies, dealing with feature extraction from gas
5 ml of pooled human blood. Blood cultured bottles were incu- sensor response [7], [8], it has been demonstrated that features
bated at 35 ◦ C for 24 h. The bacterial isolates used along with based on the dynamic of the signal can add information for the
the respective species codes are given in Table I. It is important discrimination task. Therefore, two different feature-extraction
to note that the viable count when the bottle is positive has not methods are considered in this study.
been established, and therefore, is not known at the time of sam- 1) Static response: The static response of a sensor is defined
pling. In this study, only one bacterial strain and one bacteria as the difference between the value that the sensor has at
species is considered at the time, as polyclonal or polymicrobial the end of the sampling phase minus the baseline value
bacteremia is rare and multibacteria strains arising from wounds for that sensor.
is beyond the scope of this study. 2) Response derivative: The response derivative is calculated
as the average of the derivative of the sensor during the
B. Sampling Procedure first 3 s of exposition of the array to the headspace.
Fig. 1 gives a graphical interpretation of these two features.
The purpose of the sampling system is to transfer the The extraction of these features generates a 22-D feature vector
headspace from the sample to the sensors without altering its in case only the static response is considered, or 44-D vec-
composition and properties. The sampling system of the NST tor, if both the static response and the response derivative are
3220 Emission Analyzer, Applied Sensors, Linköping, Sweden, considered.
uses two adjacent needles, one “in” needle and one “out” nee-
dle. The sample gas drawn through the “in” needle is passively
replaced with air through the “out” needle by the negative pres- B. Dimensionality Reduction
sure created in the sample bottle, as the headspace is removed. When the dimensionality of the considered feature space is
This ensures a minimal dilution factor, and therefore, a mini- high and a finite number of samples is available, a good estima-
mal alteration of the properties, in particular when compared to tion of the discriminant function is difficult to obtain. However,
systems that use a carrier gas. in many practical applications, multivariate data in n usually
The sensor array of the NST Emission Analyzer is composed have an intrinsic dimensionality much lower than n. This is es-
of 10 MOS and 12 MOSFET sensors, for a total of 22 sensors. pecially true for gas-sensing applications, where the gas sensors
These two technologies complement each other enabling the in an array have a highly correlated response. As a result, all
array to discriminate between a much wider range of gases than the samples can be enclosed in a hyperspace of dimensionality
by using just one sensor technology [6]. much lower than the one of the original feature space. There are
2886 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 57, NO. 12, DECEMBER 2010

in the one-versus-the-rest approach, one SVM for every class

Ck is trained in which samples that belong to Ck are considered
the positive class, while samples belonging to the other K − 1
classes are considered the negative class. In the one-versus-one
approach K(K − 1)/2, binary SVMs are trained for all the
possible pairs of classes. Following the indications presented
in [11], we chose to use the one-versus-one approach.

D. Posterior Probability Estimation

The SVM is a decision machine, and therefore, it does not
provide any estimation of the posterior probability P (C|x) of
sample x belonging to class C. Though many methods have been
proposed in the literature for obtaining such estimation [12],
the one considered in this paper consists in fitting a sigmoid
to every pairwise decision in order to get an estimate of the
Fig. 1. Graphical interpretation of the two feature extraction methods used in pairwise-coupled posterior probability and then ensembling the
this study. pairwise-coupled posteriors in order to get a multiclass posterior
probability. Given a two-class dataset, where the two classes
are encoded by the values y = ±1 and f is the (unthresholded)
fundamentally two approaches for reducing the dimensionality output of the SVM, a parametric form of a sigmoid that estimated
of a feature space: selecting a subset of the original feature set the posterior probability for class 1 is
(i.e., feature selection) or projecting the original feature space
into a lower dimensional one. In this study, we considered a 1
r+1−1 = P (y = +1|f ) = . (1)
method belonging to the second family, the linear discriminant 1 + eA f +B
analysis (LDA). The LDA finds a projection of the data on a The parameters A and B are calculated by minimizing the
lower dimensional space that maximizes a class separability negative log likelihood of the data
criterium based on the concepts of inter-class and intra-class  
scatter. In this work we have ten classes and the original feature 
space is projected into a 9-D space. argmin − ti log(Pi ) + (1 − ti ) log(1 − Pi ) (2)
A ,B i
Feature selection should be considered in order to be able
to optimize the sensor array. This will be considered in future yi + 1
where ti = . (3)
studies, where the development of a tailor-made instrument will 2
be addressed. More details about how the optimization problem is solved
can be found in [13]. It is important to notice that, as Platt
C. Classification pointed out in [14], the SVM decision values f are usually
clustered around ±1, and therefore, the probability estimation
The classification algorithm that has been considered in this
given by (1) might be inaccurate due to overfitting. In order
study is the SVM [9]. The SVM is a popular kernel-based al-
to limit this overfitting, the parameters A and B are estimated
gorithm that projects the data into a high-dimensional space
performing a further fivefold cross-validation on the dataset.
in which the problem is solved using a maximum margin
Once all the set of pairwise posteriors have been obtained,
linear classifier. The linear decision boundaries in the high-
they are ensembled according to the second approach proposed
dimensional feature space are in general nonlinear decision
in [12]. This method is formulated as an optimization problem
boundaries in the original feature space. One of the most im-
as follows:
portant properties of SVMs is that the estimation of the model
parameters is a convex optimization problem, and therefore, any 
k 

local solution is also a global optimum. Many variations of the minimize (rj i pi − rij pj )2 (4)
p
original model of SVM have been proposed, both for classifi- i=1 j :j = i

cation and regression problems. The model used in this study 

k
is the soft-margin SVM with Gaussian kernel. This particular subject to pi = 1 ∀i
model has two hyperparameters that need to be set in advance: i=1
γ that is the size of the Gaussian kernel, and the regularization pi ≥ 0 ∀i (5)
parameter C that determines how much samples that fall on the
wrong side of the decision boundary have to be penalized. where rab is the pairwise posterior probability calculated with
The SVM is by definition a binary classifier, though it is (1) for the two-class case, pa is the multiclass posterior proba-
possible to extend it to the multiclass case using different ap- bility for class a, and p is the vector of all the pa . It is shown
proaches. The most popular are the one-versus-the-rest and the in [12] how this problem has a unique solution and can be solved
one-versus-one approach [10]. Given the number of classes K, by a simple linear system.
TRINCAVELLI et al.: DIRECT IDENTIFICATION OF BACTERIA IN BLOOD CULTURE SAMPLES USING ELECTRONIC NOSE 2887

E. Ensembling Decisions TABLE II

CONFUSION MATRIX OBTAINED FOR THE WHOLE DATASET BY THE
As described in Section II-B, during a measurement, the sam- ALGORITHM THAT RELIES ONLY ON THE RESPONSE OF THE SENSORS (ROWS
pling cycle is repeated ten times. The pattern recognition algo- ARE THE TRUE CLASS AND COLUMNS ARE THE ESTIMATED CLASS)
rithm is applied independently to the ten samples, and therefore,
ten estimates P (C|xi ) are obtained, where i is the number of
the sample. A step in order to decrease the uncertainty in the
identification is to try to ensemble the information coming from
the ten sampling cycles instead of considering them separately.
A way to do this is to consider the ten estimates of P (C|xi ) as
a random sample, and perform a multiple comparison among
the means of P (Ck |x) for all the classes k. In this study, the
multiple comparison is performed using the Tukey’s honestly
significant differences test (HSD) [15]. It is important to notice
that a multiple comparison procedure is not equivalent to a series
of pairwise comparison, since it is designed to provide an upper
bound (the chosen confidence level α) on the probability that TABLE III
any comparison will be incorrectly found significant. Instead of CONFUSION MATRIX OBTAINED FOR THE WHOLE DATASET BY THE
performing multiple pairwise test, the confidence level α would ALGORITHM THAT RELIES BOTH ON THE RESPONSE AND ON THE DERIVATIVE
OF THE SENSORS (ROWS ARE THE TRUE CLASS AND COLUMNS ARE THE
apply to each single comparison; therefore, the chance of incor- ESTIMATED CLASS)
rectly finding a significant difference would increase with the
number of comparisons.
In this study, we use the HSD with a confidence level
α = 0.05, and we perform a classification only if, according
to the HSD result, the mean of the posterior probability for one
class k is significantly superior of the means of all the other
classes. If this does not happen, we declare a rejection for that
measurement.

IV. RESULTS
A 12-fold cross-validation on the collected dataset has been
performed to evaluate the proposed pattern recognition algo-
rithm. In every fold, one sequence of measurements have been
left out and used for testing the algorithm trained with the re-
maining 11 sequences. The hyperparameters of the SVM have
been estimated with an exhaustive grid search in the interval
[−10, 10] with step 1 in the base 2 logarithmic scale for both C
and γ. An eightfold cross-validation, where the folds have been
extracted randomly from the training set, has been performed at
every point in the grid. The confusion matrices obtained using
only the features based on the response of the sensor and using
the features based both on the response and the derivative are
displayed, respectively, in Tables II and III. The classification
accuracies obtained in the two cases are 91.8% ± 11.5% and
94% ± 12.7%, respectively. Before ensembling the posterior
probabilities calculated for the sampling cycles, an analysis of
reliability of the estimate of the posterior probabilities as a con-
fidence measure for the classification algorithm is made, as well
as an analysis of the distribution of errors with respect to the Fig. 2. Performance of the classification algorithm with a varying rejection
threshold. The upper figure shows the error rate and the lower figure the rejection
measuring cycles and measuring sequences. In order to check rate. The dashed lines represent the performance obtained by the algorithm that
the validity of the posteriors as a confidence measure, a thresh- uses only the response-based features, while the solid lines represent the perfor-
old is introduced so that, if not exceeded by the maximum of mance obtained by the algorithm that uses both the response and derivative-based
features.
the posteriors, a rejection is declared. Fig. 2 shows the results of
varying this threshold in the range (0,1) for both the considered
feature sets (response and response + derivative). The fact that dence measure for the classification algorithm. Moreover it can
the error rate decreases when the rejection threshold increases also be observed how the addition of the derivative-based fea-
indicates that the estimation of the posterior is a reliable confi- tures diminish the error of roughly the 25% over all the rejection
2888 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 57, NO. 12, DECEMBER 2010

TABLE IV
CLASSIFICATION ACCURACY FOR 12 MEASUREMENT SESSIONS

Fig. 4. Performance of the classification algorithm with a varying rejection

threshold for the dataset, where sequences 1 and 7 have been removed. The
upper figure shows the error rate and the lower figure the rejection rate. The
dashed lines represent the performance obtained by the algorithm that uses only
the response-based features, while the solid lines represent the performance
obtained by the algorithm that uses both the response and derivative-based
features.

Fig. 3. Number of errors committed by the classification algorithm in the

different measuring cycles. It is evident how the first measurement cycle is
more subject to erroneous decisions.

threshold spectrum without increasing the rejection rate. This Fig. 5. Number of errors committed by the classification algorithm in the
confirms that the dynamic characteristics of the signal contain different measuring cycles for the dataset, where sequences 1 and 7 have been
useful information for the discrimination of odors. removed. It is evident how the first measurement cycle is more subject to
erroneous decisions.
A second aspect to analyze is how errors are spread across
measurement sessions and sampling cycles. Table IV shows the
performances obtained in 12 measurement sessions. It is evident some leftover from the earlier sampling cycle is still there. This
how measurement sessions 1 and 7 obtain a performance much effect is particularly evident in the first measuring cycle, since
worse than the other sessions. This can be explained by the fact the bacteria that was smelled in the sampling cycle before was
that these two sessions are the ones recorded in the beginning different. Indeed, it is possible to use a sample, which will better
of the two experiment batches. Therefore, we can suppose that “condition” the sensors to the bacteria infected blood, which
this degradation of performance can be due to interference in is a common practice in the electronic nose community. The
the measuring system, like humidity deposited on the sensors’ effect of using a conditioner may result in a better classification
surface, the sensors were not fully warmed, or stagnant air was performance for the first sample and this will be investigated in
present in the sampling system. An analysis of how the errors future studies.
are spread across measurement cycles is given in Fig. 3. It can If the data collected during measurement sessions 1 and 7
be observed how the number of errors made during the first are removed from the set, we obtain 96.4% ± 4.3% classifi-
measuring cycle is larger than the errors in the other cycles. cation accuracy with the feature based on the static response
This can be due to the fact that the purging procedure of the and 98.9% ± 1.4% with the features that include the dynamic
nose at the end of a measuring cycle is not perfect, and therefore, of the sensors. Fig. 4 shows the effect of the introduction of a
TRINCAVELLI et al.: DIRECT IDENTIFICATION OF BACTERIA IN BLOOD CULTURE SAMPLES USING ELECTRONIC NOSE 2889

the ensemble for the dataset without measurement sessions 1

and 7 is displayed. After only four sampling cycles, perfect
discrimination is obtained, as both the error and rejection rate
are zero.
V. CONCLUSION
This paper advocates the use of electronic noses to discrim-
inate among various bacteria regularly found in the blood cul-
tures. This is an important application of electronic olfaction
that could significantly improve the current methodologies and
be successfully used in clinical settings. The results presented
show that the bacteria can be accurately discriminated using the
method. Further the proposed methods have been tested on a
large dataset, (an order of magnitude larger than earlier stud-
Fig. 6. Performance of the ensembled classification algorithm with a varying
ies). Our next step will be the starting of clinical trials to test
number of measuring cycles. The upper figure shows the error rate and the lower the robustness of the method and its applicability in a clinical
figure the rejection rate. The solid lines represents the performance obtained by setting. In particular, we will examine the effect of the geneal-
the algorithm that uses all the measuring cycles, while the dashed lines represent
the performance obtained by the algorithm that neglects the first measurement
ogy of the bacteria (i.e., different strains of the same species on
cycle. Notice that the solid line starts from cycle two and the dashed line from discrimination performance).
cycle three. This is because at least two samples are needed to calculate an
uncertainty.
ACKNOWLEDGMENT
The author would like to thank Å. Öström at Örebro Uni-
versity for providing access to the instrumentation used in this
study. The authors would also like to thank L. Barkman at Örebro
University Hospital for her assistance in the sample preparation
and her support.

REFERENCES
[1] S. Y. Lai, O. F. Deffenderfer, W. Hanson, M. P. Phillips, and E. R. Thaler,
“Identification of upper respiratory bacterial pathogens with the electronic
nose,” Laryngoscope, vol. 112, pp. 975–979, Jun. 2002.
[2] S. Aathithan, J. C. Plant, A. N. Chaudry, and G. L. French, “Diagnosis
of bacteriuria by detection of volatile organic compounds in urine us-
ing an automated headspace analyzer with multiple conducting polymer
sensors,” J. Clin. Microbiol., vol. 39, pp. 2590–2593, Jul. 2001.
[3] M. Holmberg, F. Gustafsson, E. G. Hörnsten, F. Winquist, L. E. Nilsson,
L. Ljung, and I. Lundström, “Feature extraction from sensor data on
bacterial growth,” Biotechnol. Tech., vol. 12, no. 4, pp. 319–324, 2004.
Fig. 7. Performance of the ensembled classification algorithm with a varying [4] G. C. Green, A. D. Chan, and R. A. Goubran, “An investigation into the
number of measuring cycles for the dataset where sequences 1 and 7 have been suitability of using three electronic nose instruments for the detection and
removed. Only the rejection rate is shown since the error is constantly zero. The discrimination of bacteria types,” in Proc. IEEE Eng. Med. Biol. Soc.,
solid lines represents the performance obtained by the algorithm that uses all the 2006, vol. 1, pp. 1850–1853.
measuring cycles, while the dashed lines represent the performance obtained [5] M. Bruins, A. Bos, P. L. Petit, K. Eadie, A. Rog, R. Bos, G. H. van
by the algorithm that neglects the first measurement cycle. Notice that the solid Ramshorst, and A. van Belkum, “Device-independent, real-time identifi-
line starts from cycle two and the dashed line from cycle three. This is because cation of bacterial pathogens with a metal oxide-based olfactory sensor,”
at least two samples are needed to calculate an uncertainty. Eur. J. Clin. Microbiol. Infect. Dis., vol. 28, pp. 775–780, Jul. 2009.
[6] K. Arshak, E. Moore, G. Lyons, J. Harris, and S. Clifford, “A review of
gas sensors employed in electronic nose applications,” Sens. Rev., vol. 24,
threshold on the posterior probability on this reduced dataset. no. 2, pp. 181–198, Oct. 2004.
The observations made for the full dataset are confirmed. and [7] T. Eklöv, P. Martensson, and I. Lundström, “Enhanced selectivity of mos-
fet gas sensors by systematical analysis of transient parameters,” Anal.
in this case, the improvement obtained with the introduction of Chim. Acta, vol. 353, no. 2–3, pp. 291–300, 1997.
the features based on the derivative of the signal is even more [8] R. Gutierrez-Osuna, H. T. Nagle, and S. S. Schiffman, “Transient response
significant. Fig. 5 shows that most of the errors are done in the analysis of an electronic nose using multi-exponential models,” Sens.
Actuators B, Chem., vol. 61, no. 1–3, pp. 170–182, 1999.
first measuring cycle, also when measurement sessions 1 and 7 [9] C. J. C. Burges, “A tutorial on support vector machines for pattern recog-
are left out. nition,” Data Mining Knowl. Discov., vol. 2, no. 2, pp. 121–167, 1998.
Results from ensembling the decisions for the full dataset and [10] C. M. Bishop, Pattern Recognition and Machine Learning (Information
Science and Statistics). New York: Springer-Verlag, Aug. 2006.
for the dataset without sequences 1 and 7 are shown in Figs. 6 and [11] C.-W. Hsu and C.-J. Lin, “A comparison of methods for multi-class support
7, respectively. It is important to notice how neglecting the first vector machines,” IEEE Trans. Neural Netw., vol. 13, no. 2, pp. 415–425,
cycle improves the performance of the ensemble. This confirms Mar. 2002.
[12] T.-F. Wu, C.-J. Lin, and R. C. Weng, “Probability estimates for multi-
that the first cycle contains additional noise with respect to the class classification by pairwise coupling,” J. Mach. Learning Res., vol. 5,
subsequent cycles. In particular in Fig. 7, the performance of pp. 975–1005, 2004.
2890 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 57, NO. 12, DECEMBER 2010

[13] H. T. Lin, C. J. Lin, and R. C. Weng, “A note on platt’s probabilistic outputs Amy Loutfi received the B.Sc. degree in electrical
for support vector machines,” Mach. Learn., vol. 68, no. 3, pp. 267–276, engineering from the University of New Brunswick,
Oct. 2007. Saint John, NB, Canada, and the Ph.D. degree in
[14] J. C. Platt, “Probabilistic outputs for support vector machines and com- computer science from Örebro University, Örebro,
parisons to regularized likelihood methods,” in Advances in Large Margin Sweden.
Classifiers. Cambridge, MA: MIT Press, 1999, pp. 61–74. She is a Postdoctoral Researcher at Örebro Uni-
[15] Y. Hochberg and A. C. Tamhane, Multiple Comparison Procedures (Wiley versity. She has several years experience in electronic
Series in Probability and Statistics). New York: Wiley, Nov. 1987. nose devices. Her current research interests include
the integration of electronic nose into intelligent sys-
tems, such as mobile robots, mobile robotics for en-
vironmental monitoring, artificial intelligence, and
distributed robotic systems.

Marco Trincavelli (M’08) was born on March 24,

1981, in Italy. He received the B.Sc. and M.Sc.
degrees in computer science engineering from the
Politecnico di Milano, Milan, Italy, in 2003 and
2006, respectively, and the M.Sc. degree in electri-
cal engineering and computer science from the Lund Bo Söderquist received the Ph.D. degree in clinical microbiology and infectious
Tekniska Högskola, Lund, Sweden, in 2006. Since diseases from Karolinska institutet, Stockholm, Sweden, in 1995.
spring 2006, he has been a Graduate Student in com- He is currently an Infectious Disease Physician and Clinical Microbiologist
puter science at the Center for Applied Autonomous at Örebro University Hospital, Örebro, Sweden, and an Associate Professor
Sensor Systems, ÖRebro University, Örebro, at Örebro University, Örebro. His research interests include staphylococci and
Sweden. staphylococcal infections focusing on molecular epidemiology, pathogenesis,
During 2005 and 2006, he was a Consultant at Digitec srl. Lecco, Italy, virulence, and antimicrobial resistance.
involved in the research on computer vision. His research interests include the
study of odor discrimination with an array of gas sensors.

Per Thunberg received the M.Sc. degree in engineer-

ing physics from Umeå University, Umeå, Sweden,
Silvia Coradeschi received the M.Sc. degree in phi- and the Ph.D. degree in biomedical engineering from
losophy from the University of Florence, Florence, Linköping University, Linköping, Sweden.
Italy, the M.Sc. degree in computer science from He is currently an Associate Professor at Örebro
the University of Pisa, Pisa, Italy, and the Ph.D. de- University, Örebro, Sweden, and is the Director of the
gree in computer science from Linköping University, Center of Biomedical Engineering Research, Örebro
Linköping, Sweden. University Hospital, Örebro, where he is also a MR
She is a Full Professor of information technology Physicist in the Department of Medical Physics.
at Örebro University, Örebro, Sweden. Her current
research interests include artificial olfaction, intelli-
gent systems, and anchoring, that is, establishing the
connection between the symbols used to perform ab-
stract reasoning and the physical entities, which these symbols refer to.