sensors
Article
Low-Cost Electronic Nose for the Determination of
Urinary Infections
Alba de la Rica-Martinez 1 , Gemma Martínez-Muñoz 2 , Marta Amoros Sanjuan 1 , Agustín Conesa-Celdrán 2 ,
Lucía Garcia-Moreno 1 , Gabriel Estan-Cerezo 1 , Martin J. Oates 2 , Nieves Gonzalo-Jimenez 1
and Antonio Ruiz-Canales 2, *
Citation: de la Rica-Martinez, A.;
Martínez-Muñoz, G.; Sanjuan, M.A.;
Conesa-Celdrán, A.; Garcia-Moreno,
L.; Estan-Cerezo, G.; Oates, M.J.;
Gonzalo-Jimenez, N.; Ruiz-Canales, A.
Low-Cost Electronic Nose for the
Determination of Urinary Infections.
Sensors 2024, 24, 157. https://
doi.org/10.3390/s24010157
Academic Editors: Jun Wang and
Nicole Jaffrezic-Renault
Received: 22 September 2023
Revised: 29 November 2023
Accepted: 21 December 2023
Published: 27 December 2023
Copyright: © 2023 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Sensors 2024, 24, 157. https://doi.org/10.3390/s24010157
1 Servicio de Microbiología, Hospital General Universitario de Elche, 03202 Elche, Spain;
albadelarica@gmail.com (A.d.l.R.-M.); marta.as08@gmail.com (M.A.S.);
luciaizpi.garciamoreno@gmail.com (L.G.-M.); gestan@umh.es (G.E.-C.); gonzalo_nie@gva.es (N.G.-J.)
2 Engineering Department, Miguel Hernández University of Elche, 03312 Orihuela, Spain;
agusconesaceldran@gmail.com (A.C.-C.); moates@btinternet.com (M.J.O.)
* Correspondence: acanales@umh.es
Abstract: Currently, urine samples for bacterial or fungal infections require a long diagnostic period
(48 h). In the present work, a point-of-care device known as an electronic nose (eNose) has been
designed based on the “smell print” of infections, since each one emits various volatile organic
compounds (VOC) that can be registered by the electronic systems of the device and recognized in a
very short time. Urine samples were analyzed in parallel using urine culture and eNose technology.
A total of 203 urine samples were analyzed, of which 106 were infected and 97 were not infected. A
principal component analysis (PCA) was performed using these data. The algorithm was initially
capable of correctly classifying 49% of the total samples. By using SVM-based models, it is possible
to improve the accuracy of the classification up to 74% when randomly using 85% of the data for
training and 15% for validation. The model is evaluated as having a correct classification rate of 74%.
In conclusion, the diagnostic accuracy of the eNose in urine samples is high, promising and amenable
for further improvement, and the eNose has the potential to become a feasible, reproducible, low-cost
and high-precision device to be applied in clinical practice for the diagnosis of urinary tract infections.
Keywords: urine analysis; point-of-care; electronic nose
1. Introduction
Urinary tract infections (UTIs) are one of the most frequent infections in the world, es-
pecially in women [1,2]. UTIs refer to the presence of microbial pathogens within the urinary
tract and are usually classified by site of infection (bladder (cystitis), kidney (pyelonephritis)
or urine (bacteriuria)). These infections can be asymptomatic or symptomatic, characterized
by a broad spectrum of symptoms ranging from mild irritative urination to bacteremia,
sepsis or even death [3].
According to the latest study on the prevalence of nosocomial infection in Spain
(EPINE-EPPS 2022), in its report on the prevalence of healthcare-associated infections
(HAIs), UTIs represent the third cause of infection in primary care (17.4%), after COVID-19
(23.23%) and respiratory infections (22.67%), and the second cause of HAIs (16.75%) after
surgical infections (21.29%) [4]. The incidence of UTIs varies with sex and age. From
3 months of age and up to 50–65 years, women have UTIs much more frequently than men,
and it is estimated that 50–60% of them will suffer from such infections throughout their
lives [5]. Uncomplicated acute cystitis in young women is the most prevalent infection, with
peak incidence in the years of maximum sexual activity. From the age of 65, the incidence
increases in both sexes, more markedly in men, for whom it coincides with prostate disease.
Moreover, 10.1% of cases of nosocomial bacteremia are secondary to UTIs, being the second
focus in nosocomial bacteremia after microbiologically confirmed central vascular catheter
https://www.mdpi.com/journal/sensors
Sensors 2024, 24, 157 2 of 16

(24.5%). In addition, 28.8% of community bacteremia cases are also related to UTIs, being
the first focus in primary care patients [6].
The economic implications of UTIs are enormous because of their high incidence. Costs
include doctor visits, antimicrobial prescriptions and hospitalization expenses, as well as
non-medical costs associated with travel, illness and morbidity. The estimated annual cost
of community-acquired UTIs in the United States is estimated to be approximately USD
1.6 billion, and the estimated annual cost of nosocomial UTIs in the United States ranges
from USD 424 million to USD 451 million [3].
The etiology of UTIs depends on many factors, such as age, sex, underlying conditions,
the presence of functional or anatomical disorders of the urinary tract, the community or
nosocomial origin of the UTI, the use of prolonged or intermittent catheterization, history
of recent hospitalization, institutionalization or previous antibiotic treatments.
Escherichia coli is the bacterium that is most commonly associated with UTIs. Other
bacteria that can cause UTIs include Staphylococcus saprophyticus, Proteus mirabilis, Kleb-
siella pneumoniae and Enterococcus spp. A urine culture and an antibiogram of isolated
uropathogens are essential to optimize antimicrobial treatment [6]. The required time to
have a proper diagnosis can vary from 24 to 72 h between sample collection and pathogen
identification.
There is a clear need for new, rapid, sensitive and reliable analytical methods, prefer-
ably with low operating costs, that can allow for the early detection of urinary tract infec-
tions and other diseases in urine. A key benefit would be the early selection of appropriate
antimicrobial treatments.
Mass spectrometry and the electronic nose (eNose) are two different analytical tech-
niques used for chemical analysis, particularly in the identification of volatile compounds
or odors. They serve different purposes and have their own advantages and limitations.
Mass spectrometry is a highly versatile analytical technique used to determine the
chemical composition of a sample by measuring the mass-to-charge ratio of its ions. It
is capable of providing detailed information about the chemical structure of compounds,
making it suitable for the identification of specific molecules and their quantification.
Moreover, it is highly sensitive and can detect a wide range of compounds, including
volatile and non-volatile ones. This technique is widely used in various applications,
including chemistry, biology, environmental analysis and food safety.
An electronic nose, or eNose, is an instrument that mimics the human sense of smell to
detect and identify odors and volatile compounds in the air. An eNose consists of an array
of sensors that respond to different chemical compounds. The pattern of sensor responses
is then used to identify or classify the odor or mixture of volatile compounds. eNoses are
generally less specific and less precise compared to mass spectrometry and are therefore
better suited for identifying broad classes of odors rather than specific molecules.
When considering the choice between mass spectrometry and eNoses for a specific
application, several aspects should be taken into account. According to specificity, mass
spectrometry provides high specificity and can identify individual compounds, while
eNoses provide more pattern-based responses and are better suited to identifying broad cat-
egories of odors or volatiles. Concerning sensitivity, mass spectrometry is highly sensitive
and can detect trace amounts of compounds, whereas eNoses may have lower sensitivity.
Taking into account cost and complexity, mass spectrometry equipment is typically more
expensive and requires a higher level of expertise to operate compared to eNoses.
Considering the specific requirements of the final objective, the above are the main
remarkable differences between these two devices. If it is necessary to identify and quantify
specific compounds with high precision, mass spectrometry is often the better choice. If a
quick and relatively inexpensive method to detect or classify odors is needed, an eNose
may be more appropriate.
In some cases, a combination of these techniques may be used, where an eNose is used
for initial screening, and mass spectrometry is employed for more detailed analysis when
specific compounds need to be identified [7].
Sensors 2024, 24, 157 3 of 16

The development of “electronic noses” (eNoses) in recent years has focused on the
quality control of food products. As a result, the fragrance and food/beverage industries (in
which human sensory panels are often used) have already begun to benefit from artificial
nose technology. In addition, other applications have been developed for the detection
of explosives [8], but the main field for these devices is the biomedical diagnosis for
gastrointestinal diseases [9] or even cancer [10]. Artificial smelling may have a considerable
impact on medicine in the near future; for example, research is underway in areas such as
breath or urine analysis [11].
The electronic nose plays a fundamental role in relation to the characterization of
samples according to their aroma, being an alternative to conventional sensory methods,
as well as gas chromatography techniques. eNose technology is based on a series of non-
specific sensors which react to gases and generate different signals. These signals can be
processed and used for compound identification and sample classification depending on
the emitted aromas. These aromas are associated with the chemical detection of complex
gas mixtures, which consist mainly of volatile organic compounds (VOCs). Electronic
noses are capable of detecting complex mixtures of volatile compounds present in gaseous
samples. These mixtures generate a combined response in the sensors and create an odor
pattern. The use of data analysis, such as principal component analysis (PCA), cluster
analysis and classification techniques such as artificial neural networks (ANN) or support
vector machines (SVM), have the potential to classify samples based on their aroma with a
proper level of accuracy (between 70 and 100%) [12–14].
Currently, numerous studies have been carried out on the application of low-cost
electronic noses in different samples of wine [15], olive oil [16], black tea [17] or palm tree
infections [18]. Moreover, these devices are very useful at an industrial level for quality
control, safety and traceability [19]. Their use in urinary samples can create a new, useful
tool in UTI diagnoses. Other authors have shown the use of this technology in urine [20]
and urinary tract infections [5,21–23]. In this study, two powerful techniques for data
analysis have been employed for the analysis and classification of the dataset. These
techniques are principal component analysis (PCA) and support vector machines (SVM).
The selection of these methods was driven by their respective strengths and their relevance
to the research objectives.
PCA is an essential dimensionality reduction technique that plays a crucial role in
feature engineering. It is used to transform high-dimensional data into lower dimensional
representations, preserving the most critical information while eliminating redundancy.
By reducing dimensionality, PCA simplifies the dataset, making it computationally more
efficient and less prone to overfitting. Additionally, PCA enhances the interpretability of
the data, facilitating a more intuitive understanding of the underlying patterns.
On the other hand, SVM is a robust and versatile machine learning algorithm used
for classification. The choice of SVM in this study is motivated by its capacity to handle
both linear and non-linear datasets effectively. SVM constructs a decision boundary that
maximizes the margin between data points of different classes, making it a powerful
tool for separating complex and overlapping data. SVM can handle small datasets with
high dimensionality effectively. It is particularly useful when the number of features is
comparable to or even exceeds the number of data points [24]. Furthermore, SVM exhibits
resilience against overfitting, which is vital for producing reliable and generalizable models.
Its ability to handle high-dimensional data and non-linear relationships makes SVM an
excellent choice for the experiment, where the underlying data structure may be complex
and multi-dimensional.
The objective of this work is focused on adjusting, optimizing and carrying out an
initial validation of a prototype of an electronic nose (eNose) device as a point-of-care
diagnostic method through systematic laboratory studies and then comparing it with the
current gold standard method for the diagnosis of urine infections (urine culture). To
achieve this aim, the following specific objectives have been developed:
Sensors 2024, 24, 157 4 of 16

- To quantify the operational limits for the eNose as a device for the detection of bacterial
or fungal urine infections;
- To validate the effectiveness and precision of the eNose in a clinical environment for
the point-of-care diagnoses of bacterial or fungal infections in urine.

2. Materials and Methods

2.1. Sample Selection
During the study, 203 urine samples were processed, the selection of which was made
via consecutive sampling of urine samples following requests for urine culture received
at the Microbiology Service of the Hospital General Universitario of Elche (HGUE), Elche,
Spain. The hospital provides coverage for the health department that encompasses the
municipalities of Elche and Santa Pola, with a total population of 163,576 inhabitants.
Samples from patients derived from primary care, emergency and inpatient care were
included. No age restrictions were applied; the average age was 61 years (age range
4–96 years), and 62.62% of patients were female.
Urine culture requests that met the collection criteria established by the Spanish Society
of Infectious Diseases and Clinical Microbiology (SEIMC) were analyzed. Sterile 100 mL
bottles or 10 mL vials with or without boric acid were used. Samples with less than 5 mL
of urine were excluded. Urine samples were analyzed in parallel using urine culture and
through eNose technology.

Urine Culture Analysis

Urine culture was performed quantitatively using 1 µL calibrated loops. Each isolated
colony corresponded to 1000 CFU/mL. The incubation was carried out at 35–37 ◦ C in
an enriched CO2 atmosphere in a blood agar base and at the same temperature under
an aerobiosis atmosphere in cystine–lactose–electrolyte-deficient agar (CLED agar). The
interpretation of the urine culture results was carried out according to the original criteria
proposed by Kass, considering as urinary infections those that presented ≥ 100,000 colony-
forming units (CFU) in the culture.

2.2. eNose Components and Performed Analysis

The same urine samples were analyzed using a prototype of a low-cost electronic nose
(eNose), which was designed by the work group of the Engineering Department of the
Miguel Hernandez University of Elche (UMH) in collaboration with the technology-based
company Telenatura EBT, S.L., located in Elche (Spain). The employed device is made out
of a simple sample delivery system, composed of a sample chamber, where the samples
were deposited, together with an air pump or fan, as well as a matrix of sensors and a
data-processing unit or microcontroller (Arduino Nano microcontroller with USB serial
connection, Figure 1).
The sensor array consists of eight MQS sensors (MQ-135, MQ-2, MQ-3, MQ-4, MQ-5,
MQ-7, MQ-8 and MQ-9), manufactured by Hanwei Electronics. Co., Ltd. (Zhengzhou,
China), which have resistors (RLs) that change their values in function of the analyzed
mixture of gases. MQSs are “Metal Oxide Semiconductor” resistive sensors. These are a type
of gas sensor commonly used to detect the presence of various gases in the surrounding
environment. These sensors are based on the principle that the electrical resistance of
a metal oxide material changes when it comes into contact with specific gases. They
are widely used in applications like air quality monitoring, industrial safety and gas
leak detection.
MQ gas sensors are electrochemical sensors, and their resistance varies when exposed
to certain gases. The sensors’ interior consists of a heater that is responsible for increasing
the internal temperature. Thanks to the increase in heat, the sensor reacts with the gases,
causing a change in the resistance value. The reason for using these sensors is their wide
range of detection of VOCs (see Table 1). This prototype has been properly used for
agro-food purposes in previous experiments [15,16,18].
Sensors 2024,24,
Sensors2024, 24,157
x FOR PEER REVIEW 5 of
of1616

Figure1.1.Image
Figure Imageofofthe
theemployed
employedeNose
eNosedevice.
device.

MQ gas sensors are electrochemical sensors, and their resistance varies when ex-
Table
posed1. Used sensors
to certain in the The
gases. eNose prototype
sensors’ and their
interior sensibilities.
consists of a heater that is responsible for
increasing the internal temperature. Thanks to the increase in heat, the sensor reacts with
N◦ Sensor Sensible to
the gases, causing a change in the resistance value. The reason for using these sensors is
their 1wide range of
MQ2
detection of LPG
VOCs(liquefied petroleum
(see Table gases,
1). This hydrogenhas
prototype andbeen
propane)
properly
2 MQ3 Alcohol
used for agro-food purposes in previous experiments [15,16,18].
3 MQ4 Methane
4 MQ5 Hydrogen and LPG
Table51. Used sensors
MQ7in the eNose prototype and their sensibilities.
Hydrogen and carbon monoxide
6 MQ8 Hydrogen
N° Sensor Sensible to
7 MQ9 Carbon monoxide and LPG
18 MQ2MQ135 LPG (liquefied
NH3 (ammonia), NOX , alcohol,hydrogen
petroleum gases, and propane)
benzene, smoke, CO2 , etc.
2 MQ3 Alcohol
3 MQ4 Methane
The entire device includes several parts: eight gas sensors (MQ7 and MQ9), an Arduino
4 MQ5 Hydrogen and LPG
Nano microcontroller and an analog circuit for controlling the heating of the sensors. The
5
sensitivity MQ7sensor can be improved
of each Hydrogen and carbon
by modulating the monoxide
voltage across the sensors
6 MQ8 Hydrogen
and varying the input voltage sinusoidally. Then, the system modulates the heating of the
sensors7 by varyingMQ9the voltage across them. Carbon monoxide and
A microprocessor LPG Nano generates
Arduino
voltage8 signals MQ135
for the sensors NH 3 (ammonia),
and NOX,responses.
measures their alcohol, benzene, smoke,
An analog COincluding
circuit, 2, etc.

a DAC and operational amplifiers, effectively controls the sensor heating. Moreover, the
voltageThe entire device
variation strategy includes
includes several
a dataparts:
sheet eight gas sensors.
for some sensors (MQ7 and MQ9),
Specifically, an Ar-
data sheets
duino
for someNano
sensorsmicrocontroller
(MQ7 and MQ9) and recommend
an analog circuit for controlling
a specific the heating
voltage switching of the
strategy (5.0sen-
V
sors.
and 1.4The
V) sensitivity
in a 60 + 90ofs each
cycle,sensor
with acan 90 sbesensor
improved by modulating
response time. the voltage across the
sensors andwork,
In this varying the input voltage
a sinusoidal voltage sinusoidally. Then, the system
variation is implemented withmodulates
a period of the128
heat-
s,
ranging from
ing of the 1.6 to 4.8
sensors by V, and 256
varying thesteps in each
voltage period.
across them.Additionally, the sample
A microprocessor chamber
Arduino is
Nano
made of glass
generates and connected
voltage signals for to theasensors
separate anddetection
measures chamber via PVC tubing
their responses. andcircuit,
An analog cable
glands. The detection chamber contains the sensor assembly, and a return
including a DAC and operational amplifiers, effectively controls the sensor heating. More- tube completes a
hermetically closed variation
over, the voltage circuit. Tostrategy
normalize MQ sensor
includes outputs,
a data sheet 50 forkΩ trimmer
some potentiometers
sensors. Specifically,
are used
data as load
sheets for resistors.
some sensors (MQ7 and MQ9) recommend a specific voltage switching
Potentiometer
strategy (5.0 V andvalues areaadjusted
1.4 V) in 60 + 90 s until
cycle,sensor
with achannels
90 s sensor give a voltage
response difference of
time.
less than 100 mV, at least half an hour after introducing the sample.
In this work, a sinusoidal voltage variation is implemented with a period of 128 s,
Impedance
ranging from 1.6characteristics
to 4.8 V, and 256 of each
stepssensor
in eachare balanced
period. to counteract
Additionally, manufacturing
the sample chamber
variability. This system is designed for precise and controlled
is made of glass and connected to a separate detection chamber via PVC tubing measurements of gases or cable
and sub-
glands. The detection chamber contains the sensor assembly, and a return tube completes
 a hermetically closed circuit. To normalize MQ sensor outputs, 50 kΩ trimmer potentiom-
eters are used as load resistors.
Potentiometer values are adjusted until sensor channels give a voltage difference of
Sensors 2024, 24, 157 6 of 16
less than 100 mV, at least half an hour after introducing the sample.
Impedance characteristics of each sensor are balanced to counteract manufacturing
variability. This system is designed for precise and controlled measurements of gases or
stances detected by the MQ7 and MQ9 sensors. The sinusoidal voltage variation, along with
substances detected by the MQ7 and MQ9 sensors. The sinusoidal voltage variation, along
normalization techniques, aims to enhance sensitivity and account for sensor variability.
with normalization techniques, aims to enhance sensitivity and account for sensor variability.
In addition, a second analysis is carried out to study the classification capacity of the
In addition, a second analysis is carried out to study the classification capacity of the
different infections present in the group of infected samples. In Figure 2, the motherboard
different infections present in the group of infected samples. In Figure 2, the motherboard
of the prototype device is illustrated.
of the prototype device is illustrated.

Figure 2. Motherboard of the device.

Figure 2. Motherboard of the device.

On the
On the other
other hand,
hand, aa calibration
calibration process
process is is needed
needed to to define
define the the parameters
parameters of of the
the
analysis, including the stabilization time (SBT) of the sensor,
analysis, including the stabilization time (SBT) of the sensor, prior to the start of the prior to the start of the ex-
periments; sensing time (ST), which represents the time that
experiments; sensing time (ST), which represents the time that the sensors will be exposedthe sensors will be exposed
to the
to the sample;
sample; cleaning
cleaning timetime (CT),
(CT), which
which represents
represents the the time
time that that elapses
elapses between
between two two
samples; total
samples; total time
time (OVT
(OVT == ST + CT), which represents the time necessary for the the analysis
analysis
of aa complete
of complete sample;
sample; and and thethe maximum
maximum numbernumber of of analyses
analyses (MNA), (MNA), which which isis estimated
estimated
accordingto
according tothe
theformula
formulaMNA MNA==(300 (300− − SBT)/OVT.
SBT)/OVT. This This last
last value is established to obtain obtain
a total time
a total time of the experiment lower than five hours, because longer times could negatively
affect the
affect the results
results of of the
the tests.
tests.
All experiments
All experiments took took place
place in
in aa clean,
clean, disinfected
disinfected environment.
environment. Before starting starting the the
analysis, the
analysis, theeNose
eNosesensors
sensorswerewereexposed
exposedtotoambient
ambientair airforfor3030min min(SBT).
(SBT).After
After this
this time,
time, a
10
a 10mLmL aliquot
aliquot ofofurine
urinewaswastaken,
taken,which
whichwas wasintroduced
introducedinto intothe thesample
sample chamber,
chamber, which which
consisted
consisted of of aa 140
140 mL mL glass
glass container.
container. The
The sample
sample chamber
chamber remained
remained hermetically
hermetically closed
closed
for 10 min (ST), during
for 10 min (ST), during which which time the sensors, contained in the lid of the container,
container, werewere
exposed
exposed to to the
the sample;
sample; contact
contact between
between thethe sensors
sensors andand the
the liquid
liquid was was avoided
avoided at at all
all times
times
to
to prevent
prevent measurement
measurement saturations.
saturations. Prior
Prior to introducing
introducing the the sample into the container,
container, it it
was
was manually
manually identified
identifiedusing usingaaspecific
specificsoftware
softwaredesigned
designedby byour ourresearch
researchgroup.
group.
Once
Once the the 1010 min
min had
had passed
passed with
with the
the sample
sample inside
inside the the jar,
jar, the
the sensors
sensors were
were again
again
exposed
exposed to ambient air, but this time only for 20 min (CT); this time was also correctly
to ambient air, but this time only for 20 min (CT); this time was also correctly
recorded
recordedin inthe
thesoftware
softwaredeveloped
developedbybythe theresearch
research group
group fromfrom Miguel
Miguel Hernández
Hernández Univer-
Uni-
sity of Elche (Fresh Air; version 1.0). Subsequently, successive
versity of Elche (Fresh Air; version 1.0). Subsequently, successive samples were reintro- samples were reintroduced
one
duced by one
one.by Software was created
one. Software to connect
was created the eNose
to connect the to a computer
eNose with at with
to a computer least at64least
bits
in order to obtain the data analyzed during the day, generate Excel
64 bits in order to obtain the data analyzed during the day, generate Excel files (.xls) con- files (.xls) containing
the data the
taining of each
dataofofthe daily
each of measurements and perform
the daily measurements and anperform
analysis an of the raw data
analysis forraw
of the the
extraction of characteristics. The software was also capable of creating a reading of numeri-
cal data on the sample that was being analyzed at that moment, as well as the ambient air,
which was stored in an Excel file automatically and simultaneously with the analysis. Once
the experiment ended, the file was saved in the directory that was initially chosen, as it was
a prerequisite for the analysis to choose its location and nomenclature. In Figure 3, the gas
 data for the extraction of characteristics. The software was also capable of creating a read-
ing of numerical data on the sample that was being analyzed at that moment, as well as
the ambient air, which was stored in an Excel file automatically and simultaneously with
Sensors 2024, 24, 157 the analysis. Once the experiment ended, the file was saved in the directory that was 7ini- of 16
tially chosen, as it was a prerequisite for the analysis to choose its location and nomencla-
ture. In Figure 3, the gas information response curves of the electronic nose are presented.
The main differences
information response between
curves ofthe
thesamples arenose
electronic observed in terms of
are presented. the
The frequency
main of thebe-
differences
waves, which is something very difficult to see visually, and this is why we use
tween the samples are observed in terms of the frequency of the waves, which is something frequency
analysis
very techniques.
difficult to see visually, and this is why we use frequency analysis techniques.

Figure 3. Cont.
 Sensors2024,
Sensors 2024, 24,
24, 157
x FOR PEER REVIEW 8 of 16
8 of 16

Figure 3.
Figure Gas information
3. Gas information response
responsecurves
curvesofofthe
theelectronic
electronicnose.
nose.(A) Blank
(A) Blanksample (air).
sample (B)(B)
(air). Uninfected
Unin-
urine sample.
fected (C) Infected
urine sample. urine urine
(C) Infected sample (microorganism:
sample Proteus
(microorganism: mirabilis).
Proteus mirabilis).

2.3. Data Analysis

2.3. Data Analysis
Feature extraction was first performed to analyze the raw data. Discrete Fourier
Feature extraction was first performed to analyze the raw data. Discrete Fourier
transform (DFT) was used to extract frequency coefficients from each of the temporal
transform (DFT) was used to extract frequency coefficients from each of the temporal sig-
signals of each sensor by using the same self-developed software mentioned above. In total,
nals of each sensor by using the same self-developed software mentioned above. In total,
80 coefficients were extracted for each sample, thus forming a characteristic vector together
80 coefficients were extracted for each sample, thus forming a characteristic vector to-
with a label that was later stored and used for classification.
gether with a label that was later stored and used for classification.
Secondly, a principal component analysis (PCA) was used to reduce the dimensionality
Secondly, a principal component analysis (PCA) was used to reduce the dimension-
of the data from 80 to 2. In this way, 2D graphs could be made, where each point represents
ality of the data from 80 to 2. In this way, 2D graphs could be made, where each point
one of the analyzed samples. With this tool, we can observe the clustering of the samples.
represents one of the analyzed samples. With this tool, we can observe the clustering of
It also allows us to select a sub-set of coefficients to be used for classifying the samples. In
the samples. It also allows us to select a sub-set of coefficients to be used for classifying
this case, we set the conserved variability as ≥95%, so we used 40 coefficients. The PCA
the samples. In this case, we set the conserved variability as ≥95%, so we used 40 coeffi-
and its graphic representation were executed through the Python® language, using the
cients. The PCA and its graphic representation were executed through the Python® lan-
sklearn and matplotlib libraries.
guage, using the sklearn and matplotlib libraries.
Third, the k-nearest neighbors (k-nn) algorithm was used, which is an approach for
Third, the k-nearest neighbors (k-nn) algorithm was used, which is an approach for
data classification that estimates the probability that a data point is a member of one group
data classification that estimates the probability that a data point is a member of one group
or another, according to the closest data point characteristics.
or another, according to the closest data point characteristics.
Finally, we proceed to the classification of the characteristic vectors extracted and
Finally, we proceed to the classification of the characteristic vectors extracted and
processed via PCA by using models based on support vector machines (SVMs) in order
processed via PCA by using models based on support vector machines (SVMs) in order
to obtain a fast data classification tool. The reason SVM was used is that the pipeline of
to obtain a fast data classification tool. The reason SVM was used is that the pipeline of
use of the eNose shown in this work is already established and implemented in previous
use of the eNose shown in this work is already established and implemented in previous
works that we have developed [25]. In these, the best results obtained for the eNose data
works that we have developed [25]. In these, the best results obtained for the eNose data
that we are dealing with have been using SVM. Also, SVM is very well studied and one of
that we are dealing with have been using SVM. Also, SVM is very well studied and one
the most used algorithms for classification. It has been widely employed for eNose specific
of the most used algorithms for classification. It has been widely employed for eNose spe-
data [26,27] and thus is conclusively applicable to our purpose. Since the goal of the work
cific data [26,27] and thus is conclusively applicable to our purpose. Since the goal of the
is to demonstrate the applicability of the eNose prototype and the pipeline in the medical
work is to demonstrate the applicability of the eNose prototype and the pipeline in the
field of urinary infections, no comparison with other models is reported. The SVM model
medical field of urinary infections, no comparison with other models is reported. The SVM
used is from sklearn library SVC C-Support Vector Classification. Parameters C and gamma
model used is from sklearn library SVC C-Support Vector Classification. Parameters C
for the model are obtained as explained in the text, and the kernel function is the radial
basis function (RBF). Again, the kernel function integrated into the pipeline was applied. A
Sensors 2024, 24, 157 9 of 16

search for the optimal values for C and gamma was performed to generate and train the
models. Through a recursive code using 5-fold testing and 20% test values, values for C
and gamma were obtained with a score of 0.6. The data were divided into training and
testing groups, with proportions of 85% and 15%, respectively. For the binary classification
between infected and non-infected samples, no specific classifier was selected. An onevsone
(ovo) multi-classifier was used for classification among the four types of infection.
For the analysis of the results, the following confusion matrices were made for each of
the models through which the total precision of the models, their sensitivity and specificity
were calculated:
Accuracy = the number of hits/total number of predictions;
Sensitivity = the number of positive hits/(number of positive predictions);
Specificity = the number of negative hits/(number of negative predictions).
Combining the binary classification between infected and non-infected samples and
the classification by infection, a joint model was obtained for the automatic detection of the
urine infection and type of infection.
This study was approved by the ethics committee of the University General Hospital
of Elche with the code PI 108/2021.

3. Results
3.1. Differentiation of Infected and Uninfected Urine
The starting point was 203 total samples, of which 108 were found to be infected (51.5%
of the total) and 95 to be non-infected via urine culture. Moreover, 21 different infections
were identified, of which the most common were E. coli (23 samples), Klebsiella pneumonia
(18 samples), Enterococcus faecalis (18 samples) and Proteus mirabilis (10 samples).
Figure 4 shows a graph where the first two main coefficients (PCs) that carry 69% of
the variability in the data are represented. There is no clear visual differentiation between
the control group (yellow) and the infected one (purple). The PC-based k-means algorithm
is able to discern both groups with 51.24% accuracy and classifies the entire set, which
has been entered in an unreinforced way; that is, it is capable of classifying 104 of the
Sensors 2024, 24, x FOR PEER REVIEW
203
10 of 16
samples correctly. In addition, visually, we can observe that within the groups, the infected
group shows a greater variability between samples than the control group.

Figure 4.
4. Representation
Representationofofthe
themain
maincoefficients
coefficients(PCs): control
(PCs): group
control (yellow)
group andand
(yellow) infected (purple).
infected (purple).
Sensors 2024, 24, 157 10 of 16

It is possible to improve the accuracy of the classification, with an increase in accuracy

of up to 74.19%, using SVM-based models as compared to PCA. Those models have been
trained by randomly using 85% of the data for training and 15% for validation. Thus,
the model was evaluated using 31 samples, of which 23 were properly classified, as the
confusion matrix shows in Figure 5. The higher the number of evaluated samples is, the
more reliable the obtained results will be. A higher precision when classifying non-infected
samples is observed: 82% of the non-infected samples were classified correctly compared
to 72% of the infected ones. This may be due to the biased training of the models, but there
Figure
may 4. Representation
also of themedical
be an underlying main coefficients (PCs): control group (yellow) and infected (purple).
relationship.

Figure 5.
Figure 5. Confusion
Confusionmatrix
matrixafter
afterthe application
the of of
application SVM-based models
SVM-based for the
models for control and infected
the control and in-
variables.
fected variables.

In the graph by
3.2. Sub-Analysis illustrating the PCA analysis
Group of Etiological Agents in Figure 4, it can be seen that the infected
samplesFigure 6 displays a graph that depicts the
are grouped more compactly than the uninfected ones.
results of the PCAThis is because
analysis these data
on various sam-
are more similar to each other; that is, they present less variability. On
ples of infections. The graph uses different colors to distinguish between four the other hand,
typesit is
of
observed that the infected samples are much more dispersed in the graph, which
samples, as described in the figure legend. For E. coli (in the purple color), a diagonal means
that they present much greater variability in the measurements. This could affect the model,
aggrupation distinguished from the rest can be spotted on the bottom left of the figure.
increasing its capability of classifying the infected ones by making them more linked to
No clear differentiation can be seen between Klebisiella pneumoniae (in the dark blue color)
each other as they belong to the same group.
and Enterococus faecalis (in the light blue color), which together form a diagonal clustering
above
3.2. the previous
Sub-Analysis one. Finally,
by Group when
of Etiological studying the distribution of the Proteus mirabilis
Agents
samples (in the yellow color), it is seen that they occupy a larger area of the graph and
Figure 6 displays a graph that depicts the results of the PCA analysis on various
stand apart from the other samples.
samples of infections. The graph uses different colors to distinguish between four types
of samples, as described in the figure legend. For E. coli (in the purple color), a diagonal
aggrupation distinguished from the rest can be spotted on the bottom left of the figure. No
clear differentiation can be seen between Klebisiella pneumoniae (in the dark blue color) and
Enterococus faecalis (in the light blue color), which together form a diagonal clustering above
the previous one. Finally, when studying the distribution of the Proteus mirabilis samples
(in the yellow color), it is seen that they occupy a larger area of the graph and stand apart
from the other samples.
 Sensors 2024, 24, x FOR PEER REVIEW 11 of 16
Sensors
Sensors2024,
2024,24,
24,x157
FOR PEER REVIEW 11 of 16
11 of 16

Figure 6. Representation of the main coefficients (PCs) of the most representative UTI-producing agents.
Representationofofthe
Figure6.6.Representation
Figure themain
maincoefficients
coefficients(PCs)
(PCs)ofofthe
themost
mostrepresentative
representativeUTI-producing
UTI-producingagents.
agents.

The k-nn
The k-nn algorithm
algorithm isis capable
capable ofof grouping
grouping thethe samples
samples into
into the
the four
four groups
groups of of interest
interest
withThe k-nn accuracy,
65.21% algorithm thus
is capable of grouping
classifying 45 of the samples
the 69 samplesintocorrectly
the four groups
according of interest
to their
their
with 65.21% accuracy, thus classifying 45 of the 69 samples correctly according to
with 65.21%
infection. It accuracy,
It is
is possible thus
possible to classifying
to improve
improve the 45 of the
the accuracy
accuracy of 69 samples
of the correctly
the classification
classification by according
by using to
using SVM-based their
SVM-based
infection.
infection. It is possible
models, increasing
increasing to improve
it up
up 81%. To the accuracy
To achieve
achieve of the classification
this improvement,
improvement, by using
the models
models wereSVM-based
trained byby
models, it 81%. this the were trained
models,
randomly increasing
using it up
85% of 81%.
the To achieve
data for this improvement,
training and 15% for the modelsThat
validation. were trained
is, the by
model
randomly using 85% of the data for training and 15% for validation. That is, the model
randomly usingwith
was evaluated
evaluated 85% eleven
of the data for training
samples, whichand
of which 15%
nine for classified
were validation. That is, This
correctly. the model
can be
be
was with eleven samples, of nine were classified correctly. This can
was
seenevaluated
in the with
confusion eleven
matrix samples,
shown of
in which
Figure nine
7. were
Labels classified
0–3 correctly.
correspond to E. This
coli, can be
Klebsiella
seen in the confusion matrix shown in Figure 7. Labels 0–3 correspond to E. coli, Klebsiella
seen in the confusion
pneumoniae, matrix
Enterococcus shown
faecalis andin Figuremirabilis.
Proteus 7. Labels 0–3 correspond to E. coli, Klebsiella
pneumoniae, Enterococcus faecalis and Proteus mirabilis.
pneumoniae, Enterococcus faecalis and Proteus mirabilis.

Figure 7.
Figure 7. Confusion
Confusionmatrix
matrixafter
afterapplying SVM-based
applying models
SVM-based for for
models the the
most representative
most UTI-pro-
representative UTI-
Figure
ducing7.agents.
Confusion matrix after applying SVM-based models for the most representative UTI-pro-
producing agents.
ducing agents.
Sensors 2024, 24, 157 12 of 16

4. Discussion
These results demonstrate the potential of the eNose in the detection of alterations in
urinary volatilome associated with urinary infection. By using models based on SVM, it
is possible to improve the accuracy of the classification, increasing it up to 74%, causing a
significant impact on the classification of the samples. In the present work, the accuracy of
the model increased from 51.24 to 74.00% (using eNose data).
Concerning the specific prototype device used in this paper, several improvements
can be applied. Mainly, it is possible to optimize the accuracy by selecting only the sensors
that are involved in the detection of urine infection. This aspect was not developed initially
because this device was not used previously for this purpose. Following the analysis of
the sensors involved in the detection of urine infection, it will be possible to select the best
sensors and add new specific sensors. This can improve the accuracy of the device and can
be adapted for a commercial eNose.
eNose technology offers several significant advantages in the analysis of urinary tract
infections compared to other methods. The eNose is a rapid and straightforward tool that
allows for quick analysis, which can be particularly beneficial in emergency situations. In
the analysis of urinary tract infections, accuracy has traditionally relied on techniques such
as microbiological culture (sensitivity is 95%) [28], for which the main limitation is the time
that it takes to obtain a result (24–48 h). Other methods like urinary sediment analysis
(sensitivity is 90–99%) [29] require the interpretation of values and technical equipment
that is not feasible to have in an emergency department. Compared to these conventional
methods, the eNose technique aims to provide easily interpretable results with a point-
of-care medical device suitable for integration into emergency services, thereby reducing
the required time for a diagnosis. A positive result leads to an initial treatment against the
infection, reducing it within 48 h, versus the usual diagnostic methods. A negative result
should be validated by sending a urine sample to carry out the culturing of the sample in
order to confirm the result obtained by the eNose. At least 74% of urinary tract infections
will begin the treatment within 48 h. At that point, culture tests will not be substituted by
an eNose, but they can be complemented. Another significant advantage of the eNose is
its capability to detect a wide range of volatile compounds present in urine samples. This
offers extensive coverage and enables the identification of potential infections caused by
multiple pathogens. Although an accuracy rate of 74.19% is lower than in other techniques,
we consider it high given the low economic cost and speed at which results are generated.
Furthermore, as eNose technology continues to be developed and refined, its accuracy is
likely to improve. Taking into account the aforementioned advantages and the limitations
of other available methods, we believe that eNose technology holds great potential and can
be a valuable tool in the analysis of urinary tract infections. However, as always, further
studies and validations are necessary to support and further enhance the results obtained
thus far.
In addition, it should also be noted that the model classification rate is highly depen-
dent on the sample size. For this reason, we consider it necessary to carry out more studies
to continue training the algorithm and achieve greater sensitivity and specificity.
In the graph produced by the PCA analysis in Figure 4, it can be seen that the infected
samples cluster more compactly than the uninfected ones. This can be understood as these
data being more similar to each other; that is, they present less variability. On the other
hand, it is observed that the uninfected samples are much more dispersed in the graph,
which means that they present much greater variability in their measurements. This could
affect the model, making it more capable of classifying infections by being able to link them
better as part of the same group.
The percentage of non-infected samples that were properly classified was higher (82%
instead of 72% of infected ones). This can be explained by the basis of eNose recognition
because it is based on a complex interpretation of a multi-variate non-selective response to
VOCs. There are various reasons for the limitations found in the present study. A possible
explanation could be urinary infections produced by microorganisms not detectable by
Sensors 2024, 24, 157 13 of 16

traditional cultures, such as Mycobacterium tuberculosis, Chlamydia trachomatis or the genus

Mycoplasma spp. In addition, urine metabolome is highly dependent on diet, environment
and lifestyle, and consequently exhibits greater daily variability than serum or plasma [30].
It is also known that many types of pathologies are characterized by a certain odor and
consequently, by a set of volatile substances. Many substances that are metabolites in
pathogenic processes are found in human secretions, and some authors state that ethanol,
butanol and their oxidation products (acetic and butyric acids) are present in almost all
samples [31]. Those authors, who also measured VOCs in urine with a different device,
when analyzing their results to predict the bacterial contamination of urine samples, ob-
served that the group of samples without bacteria in the urine was more dispersed, which
is similar to the observations made in our study. The authors associate this dispersion with
additional factors, such as the presence of mucus and salts in the urine, which affects the
redistribution of volatile substances at the “gas-liquid” phase boundary and thus the results
obtained through an analysis using a series of sensors. It is very likely that further research
on urine characterization will improve the correct classification of infected/uninfected
samples by removing these confounding factors.
Nowadays, in clinical practice, it is quite common to prescribe antibiotics, even when
there is no clear indication that it is a bacterial etiology. This is because there is no diagnostic
tool that can quickly determine the presence of microorganisms and their sensitivity to
various groups of antibiotics [31]. Unnecessary consumption of antibiotics leads to an
increase in antimicrobial resistance (AMR) [32], which is a major health problem recognized
by leading professional bodies around the world, including the World Health Organization
(WHO). The current estimated annual global deaths attributed to AMR is 700,000, and this
number is projected to rise rapidly and reach a worryingly high number of 10 million deaths
per year by 2050. In addition to the direct impact on human health, AMR is associated with
a significant global economic burden, with rising healthcare costs related to hospitalization
and drug use [33].
A targeted antibiotic treatment directly implies a decrease in resistance produced by
microorganisms due to inadequate treatments, as well as better therapeutic management of
the patient and also the avoidance of repeat visits to health services. Diagnostics tools that
make it possible to quickly and cheaply identify urinary tract infections will be of great
help in these aspects.
Diagnostics are a vital part of the healthcare infrastructure, with substantial impacts on
the treatment of individual patients, as well as public health policy [34]. eNose technology
meets the main requirements of this type of technique. The sensors meet a number of
criteria such as broad selectivity, high sensitivity and fast response and recovery time. In
addition, the device can be easily reproduced, is accessible, consumes less power and has
fewer consumables [20].
Despite the fact that the results for differentiating the etiological agent producing the
infection are not very robust so far, they represent a good precedent for increasing the pre-
cision of this diagnostic test. The results obtained can probably be improved by increasing
the number of samples used for system training. Therefore, further research on this topic is
necessary. In addition, the production of certain VOCs is characteristic of bacteria [11], so
it is very likely that perfecting this technology will allow us to diagnose other infectious
pathologies in the future by measuring the emissions of these compounds in different
biological fluids, such as pleural fluids, bronchoalveolar lavages or joint fluids. Research
in this field is necessary to find new lines of diagnosis that speed up the optimization of
antimicrobial treatments.
The high prevalence of urinary tract infections makes this pathology a good objective
for improving the healthcare system. The identification of new and precise diagnostic tools
capable of selecting patients who need antibiotic treatment continues to be a fundamental
issue in the healthcare field. The economic percentage that it represents for any health sys-
tem suggests that investment in advances aimed at optimizing the therapeutic management
of this disease will have a positive impact in a short time.
Sensors 2024, 24, 157 14 of 16

5. Conclusions
In closing, the diagnostic accuracy of eNoses for specific VOCs in urine samples is
high, promising and amenable for further improvement.
The presented eNose prototype has the potential to become a feasible, reproducible,
low-cost and high-precision device to be applied in clinical practice for the diagnosis of
urinary tract infections.
The application of this technology at point of care can provide great benefits at the
clinical level, improving the quality of care at the health level, with fewer unnecessary
costs, while reducing the generation of antimicrobial resistances.
The results obtained in this study highlight the need for further research to characterize
urine VOCs to improve the specificity of the technique and to obtain a sensitive and precise
device that is very useful in the pathology of urinary tract infections.
Subsequent experiments have to be developed in order to quantify and validate this
technology. Firstly, the distinction between non-infected and infected urine has been very
accurately detected. Among the infected urine samples, however, the types of infection
have not been well distinguished. This aspect has to be improved in future experiments.
Finally, the detection and quantification of the main implied metabolites (VOCs) have
to be taken into account for the appropriate validation of this technology.

Author Contributions: Conceptualization, A.d.l.R.-M. and A.R.-C.; methodology, A.C.-C. and M.J.O.;
software, A.C.-C.; validation, G.M.-M., M.A.S. and L.G.-M.; formal analysis, M.J.O.; investigation,
G.M.-M., M.A.S., L.G.-M. and A.C.-C.; resources, A.d.l.R.-M., N.G.-J. and M.J.O.; data curation,
A.C.-C. and A.R.-C.; writing—original draft preparation, A.d.l.R.-M. and G.E.-C.; writing—review
and editing, G.E.-C.; visualization, N.G.-J. and A.R.-C.; supervision, N.G.-J.; project administration,
A.d.l.R.-M. and A.R.-C.; funding acquisition, A.d.l.R.-M., G.E.-C. and A.R.-C. All authors have read
and agreed to the published version of the manuscript.
Funding: This work has been funded by the ILISABIO program from Fundación para el Fomento de
la Investigación Sanitaria y Biomédica de la Comunitat Valenciana (FISABIO) and Miguel Hernández
University of Elche (Spain): Projects DINBACT (ILISABIO22_PI03) and DINBACT (ILISABIO21_A07).
Institutional Review Board Statement: The study was conducted in accordance with the Declaration
of Helsinki, and approved by the Institutional Review Board (or Ethics Committee) of University
General Hospital of Elche (protocol code PI 108/202; November 24th 2021).
Informed Consent Statement: Patient consent was waived due to the authorization of the ethics
committee, because all samples were diagnostic surplus and they were appropriately anonymized.
Data Availability Statement: Data are contained within the article.
Acknowledgments: The authors are grateful to TELENATURA EBT, S.L. company for its technical
support in this experiment.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Tan, C.W.; Chlebicki, M.P. Urinary tract infections in adults. Singap. Med. J. 2016, 57, 485–490. [CrossRef] [PubMed]
2. Al-Badr, A.; Al-Shaikh, G. Recurrent Urinary Tract Infections Management in Women: A review. Sultan Qaboos Univ. Med. J. 2013,
13, 359–367. [CrossRef] [PubMed]
3. Foxman, B. Epidemiology of urinary tract infections: Incidence, morbidity, and economic costs. Am. J. Med. 2002, 113 (Suppl. S1),
5S–13S. [CrossRef] [PubMed]
4. ESTUDIO EPINE-EPPS nº 32: 2022. Informe España: Prevalencia de Infecciones (Relacionadas Con la Asistencia Sanitaria y
Comunitarias) y Uso de Antimicrobianos en Hospitales de Agudos. Available online: https://epine.es/api/documento-publico/
2022%20EPINE%20Informe%20Espa%C3%B1a%2020221201.pdf/reports-esp (accessed on 28 February 2023).
5. Dospinescu, V.M.; Tiele, A.; Covington, J.A. Sniffing Out Urinary Tract Infection-Diagnosis Based on Volatile Organic Compounds
and Smell Profile. Biosensors 2020, 10, 83. [CrossRef]
6. Procedimiento de Microbiología Clínica: Recomendaciones de la Sociedad Española de Enfermedades Infecciosas y Microbiología
Clínica. 14b Diagnóstico Microbiológico de Las Infecciones del Tracto Urinario. Available online: https://seimc.org/contenidos/
documentoscientificos/procedimientosmicrobiologia/seimc-procedimiento14a.pdf (accessed on 28 February 2023).
Sensors 2024, 24, 157 15 of 16

7. Ward, R.J.; Jjunju, F.P.M.; Griffith, E.J.; Wuerger, S.M.; Marshal, A. Artificial odour-vision syneasthesia via olfactory sensory
argumentation. IEEE Sens. J. 2020, 21, 6784–6792. [CrossRef]
8. Wasilewski, T.; G˛ebicki, J. Emerging strategies for enhancing detection of explosives by artificial olfaction. Microchem. J. 2021, 164,
106025. [CrossRef]
9. Wilson, A.D. Application of Electronic-Nose Technologies and VOC-Biomarkers for the Noninvasive Early Diagnosis of Gastroin-
testinal Diseases †. Sensors 2018, 18, 2613. [CrossRef]
10. Leunis, N.; Boumans, M.L.; Kremer, B.; Din, S.; Stobberingh, E.; Kessels, A.G.; Kross, K.W. Application of an electronic nose in the
diagnosis of head and neck cancer. Laryngoscope 2014, 124, 1377–1381. [CrossRef]
11. Guernion, N.; Ratcliffe, N.M.; Spencer-Phillips, P.T.; Howe, R.A. Identifying bacteria in human urine: Current practice and the
potential for rapid, near-patient diagnosis by sensing volatile organic compounds. Clin. Chem. Lab. Med. 2001, 39, 893–906.
[CrossRef]
12. Caya, M.V.C.; Maramba, R.G.; Mendoza, J.S.D.; Suman, P.S. Characterization and Classification of Coffee Bean Types using
Support Vector Machine. In Proceedings of the 2020 IEEE 12th International Conference on Humanoid, Nanotechnology,
Information Technology, Communication and Control, Environment and Management (HNICEM), Manila, Philippines, 3–7
December 2020. [CrossRef]
13. Luan, F.; Liu, H.T.; Wen, Y.Y.; Zhang, X.Y. Classification of the fragrance properties of chemical compounds based on support
vector machine and linear discriminant analysis. Flavour Fragance J. 2008, 23, 232–238. [CrossRef]
14. Murti, M.A.; Setianingsih, C.; Kusumawardhani, E.; Farhan, R. Cedarwood Quality Classification using SVM Classifier and
Convolutional Neural Network (CNN). Comput. Sci. 2022, 13, 101–111. [CrossRef]
15. Celdrán, A.C.; Oates, M.J.; Molina Cabrera, C.; Pangua, C.; Tardaguila, J.; Ruiz-Canales, A. Low-Cost Electronic Nose for Wine
Variety Identification through Machine Learning Algorithms. Agronomy 2022, 12, 2627. [CrossRef]
16. Oates, M.J.; Fox, P.; Sanchez-Rodriguez, L.; Carbonell-Barrachina, Á.A.; Ruiz-Canales, A. DFT based classification of olive oil type
using a sinusoidally heated, low cost electronic nose. Comput. Electron. Agric. 2018, 155, 348–358. [CrossRef]
17. Banerjee, M.B.; Roy, R.B.; Tudu, B.; Bandyopadhyay, R.; Bhattacharyya, N. Black tea classification employing feature fusion of
E-Nose and E-Tongue responses. J. Food Eng. 2019, 244, 55–63. [CrossRef]
18. Oates, M.J.; Abu-Khalaf, N.; Molina-Cabrera, C.; Ruiz-Canales, A.; Ramos, J.; Bahder, B.W. Detection of lethal bronzing disease in
cabbage palms (Sabal palmetto) using a low-cost electronic nose. Biosensors 2020, 10, 188. [CrossRef] [PubMed]
19. Campo Ceballos, D.A.; Mirandaeneses, J.M.; Gaviria-López, C.A.; Guerrero Narváez, J.A.; Rebelo Luna, D.A.; Hoyos García, J.
Estudio de fragancia y aroma del café tostado con la nariz electrónica Coffee-NOSE. In Proceedings of the 2020 IX International
Congress of Mechatronics Engineering and Automation (CIIMA), Cartagena, Colombia, 4–6 November 2020; IEEE: New York,
NY, USA; pp. 1–6. [CrossRef]
20. Taverna, G.; Grizzi, F.; Tidu, L.; Bax, C.; Zanoni, M.; Vota, P.; Lotesoriere, B.J.; Prudenza, S.; Magagnin, L.; Langfelder, G.; et al.
Accuracy of a new electronic nose for prostate cancer diagnosis in urine samples. Int. J. Urol. 2022, 29, 890–896. [CrossRef]
21. Kodogiannis, V.; Wadge, E. The use of gas-sensor arrays to diagnose urinary tract infections. Int. J. Neural Syst. 2005, 15, 363–376.
[CrossRef]
22. Kodogiannis, V.S.; Lygouras, J.N.; Tarczynski, A.; Chowdrey, H.S. Artificial odor discrimination system using electronic nose and
neural networks for the identification of urinary tract infection. IEEE Trans. Inf. Technol. Biomed. 2008, 12, 707–713. [CrossRef]
23. Pavlou, A.K.; Magan, N.; McNulty, C.; Jones, J.; Sharp, D.; Brown, J.; Turner, A.P. Use of an electronic nose system for diagnoses
of urinary tract infections. Biosens. Bioelectron. 2002, 17, 893–899. [CrossRef]
24. Hastie, T.; Tibshirani, R.; Friedman, J. Chapter 12. In The Elements of Statistical Learning: Data Mining, Inference, and Prediction;
Section 12.3.3 discusses SVMs and their effectiveness in high-dimensional spaces; Springer: Berlin/Heidelberg, Germany, 2009.
[CrossRef]
25. María, E.G.; Luna, A.M.; Celdrán, A.C.; Muñoz, G.M.; Oates, M.J.; Ruiz-Canales, A. Classification of Monofloral Honeys by
Measuring a Low-Cost Electronic Nose Prototype Based on Resistive Metal Oxide Sensors. Agronomy 2023, 13, 2183. [CrossRef]
26. Gaudioso, M.; Khalaf, W.; Pace, C. On the Use of the SVM Approach in Analyzing an Electronic Nose. In Proceedings of the 7th
International Conference on Hybrid Intelligent Systems (HIS 2007), Kaiserslautern, Germany, 17–19 September 2007; pp. 42–46.
[CrossRef]
27. Pardo, M.; Sberveglieri, G. Classification of electronic nose data with support vector machines. Sens. Actuators B Chem. 2005, 107,
730–737. [CrossRef]
28. Gur’ev, A.S.; Tigasson, M.; Shalatova, O.Y.; Rastopov, S.F.; Bilozor, A.; Ivanova, M.; Volkov, A.Y. Fast antibiotic susceptibility
testing of urine microflora using a microbiological analyzer based on coherent fluctuation nephelometry. Braz. J. Microbiol. 2022,
53, 195–204. [CrossRef] [PubMed]
29. Dal Canto, M.; Bartoletti, R.; Travaglini, F.; Piazzini, M.; Lodovichi, G.; Rizzo, M.; Selli, C. Molecular urinary sediment analysis in
patients with transitional cell bladder carcinoma. Anticancer. Res. 2003, 23, 5095–5100.
30. Vignoli, A.; Risi, E.; McCartney, A.; Migliaccio, I.; Moretti, E.; Malorni, L.; Luchinat, C.; Biganzoli, L.; Tenori, L. Precision Oncology
via NMR-Based Metabolomics: A Review on Breast Cancer. Int. J. Mol. Sci. 2021, 22, 4687. [CrossRef] [PubMed]
31. Kuchmenko, T.; Menzhulina, D.; Shuba, A. Noninvasive Detection of Bacterial Infection in Children Using Piezoelectric E-Nose.
Sensors 2022, 22, 8496. [CrossRef]
Sensors 2024, 24, 157 16 of 16

32. Nikolaisen, N.K.; Fertner, M.; Lassen, D.C.K.; Chehabi, C.N.; Ronaghinia, A.A.; Chriél, M.; Jensen, V.F.; Jensen, L.B.; Pedersen, K.;
Struve, T. Association between Antibiotic Consumption and Resistance in Mink Production. Antibiotics 2022, 11, 927. [CrossRef]
33. Chang, R.Y.K.; Nang, S.C.; Chan, H.K.; Li, J. Novel antimicrobial agents for combating antibiotic-resistant bacteria. Adv. Drug
Deliv. Rev. 2022, 187, 114378. [CrossRef]
34. Piorino, F.; Patterson, A.T.; Styczynski, M.P. Low-cost, point-of-care biomarker quantification. Curr. Opin. Biotechnol. 2022, 76,
102738. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.