Quantitation of ethanol in UTI assay for volatile organic compound detection by electronic nose using the validated headspace GC-MS method.

Disease detection through gas analysis has long been the topic of many studies because of its potential as a rapid diagnostic technique. In particular, the pathogens that cause urinary tract infection (UTI) have been shown to generate different profiles of volatile organic compounds, thus enabling the discrimination of causative agents using an electronic nose. While past studies have performed data collection on either agar culture or jellified urine culture, this study measures the headspace volume of liquid urine culture samples. Evaporation of the liquid and the presence of background compounds during electronic nose (e-nose) device operation could introduce variability to the collected data. Therefore, a headspace gas chromatography-mass spectrometry method was developed and validated for quantitating ethanol in the headspace of the urine samples. By leveraging the new method to characterize the sample stability during e-nose measurement, it was revealed that ethanol concentration dropped more than 15% after only three measurement cycles, which equal 30 minutes for this study. It was further shown that by using only data within the first three cycles, better accuracies for between-day classification were achieved, which was 73.7% and 97.0%, compared to using data from within the first nine cycles, which resulted in 65.0% and 81.1% accuracies. Therefore, the newly developed method provides better quality control for data collection, paving ways for the future establishment of a training data library for UTI.

Introduction

It is widely known that different types of disease can influence the pattern of volatile organic biomarkers released in exhaled gas or waste materials such as fecal matter or urine [1-3]. This is especially true for infectious diseases, stemming from the fact that different bacterial species can generate different profiles of volatile organic compounds (VOCs). Indeed, Bos et al. systematically reviewed the literature for the headspace VOC of six bacterial species and found that while they share some compounds, they form different patterns as some compounds are uniquely produced by only certain species [4]. Even among compounds that are commonly produced by multiple species, a study using Selected Ion Flow Tube Mass Spectrometry (SIFT-MS) revealed that the produced concentrations are different, thus still ensuring the unique pattern for species differentiation [5]. These fingerprint VOC profiles can be examined with headspace analysis and classified using a computer algorithm to enable remarkable detection of urinary tract infection [3, 6]. A recent review by Dospinescu et al. on studies of UTI-associated bacterial VOCs revealed a heavy reliance on Gas-liquid chromatography to discriminate VOC patterns until recently when e-noses emerged as a promising technology for identifying infectious UTI-causing strains in a clinical setting [7]. The use of the e-nose as a rapid diagnostic tool is valuable for clinical use because the gold standard of UTI detection is bacterial culture which can take days to obtain a result. While a dipstick is also a rapid, fairly accurate, and low-cost candidate, a key advantage of the e-nose is the potential to identify specific bacterial species in real-time.

Static headspace gas chromatography is a technique used for the concentration analysis of volatile organic compounds. This technique is relatively simple and can provide sensitivity that is similar to dynamic purge and trap analysis. The popularity of this technique has grown and gained worldwide acceptance for analyses of alcohol in blood, urine, and other biological samples, as well as residual solvents in pharmaceutical products. Sample matrices like blood, plastic, and cosmetics contain high molecular weights, non-volatile material that can remain in the GC system and result in poor analytical performance. Many laboratory analysts use extensive sample preparation techniques to extract and concentrate the compounds of interest from this unwanted non-volatile material. These extraction and concentration techniques can become time-consuming and costly. Static headspace analysis avoids this time and cost by directly sampling the volatile headspace from the container in which the sample is placed.

While state-of-the-art methods such as GC-MS and SIFT-MS are useful in characterizing VOC profiles, they are usually expensive, non-portable, and time-consuming in terms of sample preparation. As alternatives, there are a plethora of studies using commercially available sensors for the same purpose, which led to a technology called Electronic Nose (e-nose) [8]. An e-nose is essentially a system of gas sensors with pattern recognition capability to detect diseases. An e-nose must consist of appropriate hardware and software. Hardware refers to the type of sensors, most commercially available being Metal Oxide (MOX), and software refers to the analytical method used, usually in the form of a machine learning algorithm. Table 1 summarizes some different combinations of sensors and analytical methods used in past studies. Noticeably, a minimum of six sensors with an artificial neural network (ANN) model was enough to yield a high prediction accuracy, albeit a small number of labels of two.

Table 1

Literature survey of sensors and classification models for e-nose studies.

Study	Sensors	Analysis	Labels	Medium	Average Accuracy
Craven (1997) [9]	6 MOX	Multi-layer perceptron and Linear discriminant analysis (LDA)	4	Nutrient broth	82.2%
Gibson et al. (1997) [10]	14 Conductive polymers	Multi-layer perceptron	13	Nutrient agar	89.7%
Gardner et al. (2000) [11]	6 MOX	ANN with backpropagation	2	Blood agar then nutrient broth	96.0%
Pavlou et al. (2002) [12]	14 Conductive polymers	Genetic algorithm and backpropagation neural network	4	Agar, Brain heart infusion, cooked meat broth	95.0%
Roine et al. (2014) [6]	Commercial ion mobility spectrometer-based e-nose and 6 MOX	LDA and logistic regression	5	Normal urine made into an agar	83.9%
Asimakopoulos (2014) [13]	8 metalloporphyrin-coated sensors	Supervised Partial Least Square–Discriminant Analysis	2	Urine	84.8%
Aathithan et al. (2001) [14]	4 conductive polymers	Principal component analysis (PCA)	2	Artificial urine	72.30% sensitivity; 89.38% specificity
Seesaard et al. (2016) [15]	4 nanocomposites	PCA and cluster analysis	2	Urine	99.5%
Filianoti et al. (2022) [16]	Cyranose 320	Linear canonical discriminant analysis	2	Urine	85.3%
Seesaard et al. (2020) [17]	A hybrid of 3 nanocomposites and 3 MOX	PCA and cluster analysis	4	Bacterial culture media	99.7%
Yumang et al. (2020) [18]	7 MOX	PCA then K-nearest neighbor analysis	2	Urine	90%
Esfahani et al. (2018) [19]	Field-Asymmetric Ion Mobility Spectrometry (FAIMS) or FOX4000 (18 MOX)	Sparse Logistic Regression, Random Forest, Gaussian Process, and Support Vector	2	Urine	85–94%, depending on e-nose choice and sample age

While many more e-nose technologies exist for other applications, Table 1 is a survey of urine-based or pathogen detection studies. Many of these studies were conducted using non-physiological media such as nutrient broth, culture media, agar, or urine made into an agar. On the other hand, many urine-based studies did not involve pathogen classification in the context of UTI. Using LDA, PCA, and cluster analysis, many authors aimed to discriminate between only two labels, healthy controls and diseased samples for a variety of diseases, such as prostate cancer [13, 16], bacteriuria [14], diabetes mellitus [15, 19], and Azotemia [18]. We envisioned that point-of-care prediction of UTI should be performed directly on a patient’s urine sample with more labels to identify a wider range of causative agents. This combined lack of e-nose studies on liquid urine and a more complex classification model presents a need for urine-based tests for UTI on an e-nose equipped with a more powerful classification model.

Escherichia coli is known to be the most prevalent cause of UTI [20]. Therefore, a simple in vitro UTI model could be created by infecting commercially available human urine with pathogenic E. coli to test the e-nose system. Ethanol is a by-product of E. coli metabolism with lactose and arabinose and was considered a biomarker for E. coli, and ethanol level was also indicative of bacterial concentration in the sample [7]. However, urine samples are known to degrade over time [21, 22]. Volatile urine output has been shown to decrease after nine months of storage at -80°C [23]. Urine analysis performed on samples stored within a year resulted in much better accuracy than including samples within four years [19]. Therefore, minimization of the variability in the collected data caused by evaporation and sample degradation must be achieved before confidently and correctly establishing a library of training data for future real-time, point-of-care prediction of UTI. Therefore, a GC-MS method was needed for characterizing the headspace samples so that decisions can be made regarding the optimal measurement time, followed by the exclusion of examples that do not genuinely represent the intended labels, which could poison the analytical model.

There was little evidence in the literature that there exists an exact method for the determination of ethanol concentration in urine using headspace GC-MS. Table 2 summarizes the literature survey. Also, Tangerman argued that the headspace technique is notorious for requiring a substantial amount of labor and volumes of the biological specimen while being less sensitive [24]. However, the headspace method more closely represents the approach done with the e-noses. Therefore, a new method for the quantitation of ethanol concentration in the headspace volume of urine samples inoculated with E. coli was developed and validated in this study. Finally, it was used to help improve the classification accuracy of the EVA.

Table 2

Literature survey of gc methods for the quantitation of ethanol in biological matrices.

Study	Analyte	Sample Matrix	Method
Mihretu et al. (2020) [25]	Ethanol	Blood	Headspace GC-FID (Flame ionization detector)
Chun et al. (2016) [26]	Alcohols	Brain tissue	Headspace GC-FID
Xiao et al. (2014) [27]	Ethanol	Blood	Headspace GC-MS
Kristoffersen et al. (2006) [28]	Ethanol	Whole blood and plasma	Headspace GC-FID
Smith et al. (1999) [29]	Alcohols	Urine	Headspace SIFT-MS
Tangerman (1997) [24]	Ethanol	Whole blood, serum, urine, fecal supernatants	Direct Injection GC-MS

There are also portable systems capable of detecting and quantifying ethanol, such as zNose, or GC-IMS, which are gas sensors coupled with a GC column, or an alcohol breath analyzer, which employs a single optical gas sensor. However, we are interested in quantifying ethanol to assist with establishing a working protocol for the EVA. Therefore, a GC-MS method was developed and validated for ethanol quantification in urine. To our knowledge, this is the first paper describing the validation and quantitation of alcohol release from the bacteria by using inoculated urine samples. Using the method, we devised a plan to determine the optimal time length for measuring the urine samples to maximize the data collected without sacrificing classification accuracy.

In this paper, we use the Electronic Volatile Analyzer (EVA), a rapid, gas-sensing Internet of Things (IoT) platform under development at IBM Research, Almaden–designed to deliver classification results in under two minutes [30]. The main objective of this study was to demonstrate the utility of EVA in differentiating between a nonpathogenic (K12) and a uropathogenic (UPEC) strain of E. coli in inoculated urine and the possibility of classification improvement by using a validated GC-MS method to quantify ethanol reduction as a sign of diminished sample quality.

Materials and methods

The IBM electronic volatile analyzer™

The EVA (Fig 1) an electronic nose under development at IBM Research Almaden. The platform consists of a modular design with each gas sensor mounted on its own sensor module which is a printed circuit board of common design carrying an integrated microcontroller and the circuitry needed to operate the sensor. The sensor modules communicate via I2C protocol with a central hub (BeagleBone Black), which is a single-board computer that orchestrates the modules and processes the multi-sensorial output. Table 3 describes the set of six commercial MOX gas sensors used to collect the measurements described in the following sections. The selection of the sensor portfolio was based on the indications of their respective manufacturers regarding the nominal target gases of each sensor. It is imperative to include sensors for various compounds to capture enough differential responses for pattern recognition of complex VOC profiles. These sensors together target CO, H2, ethanol, methane, ammonia, H2S, and unspecified combustible gases. Consequently, the as-selected sensor array was expected to respond to a variety of volatile molecules, including alcohols, hydrocarbons, ammonia, methane, as well as a wide range of volatile organic compounds (VOC), with a lower limit of detection at parts per million (ppm) level.

Fig 1

The Electronic Volatile Analyzer (EVA).

(A) Sensor array prototype with parts labeled. (B) A block diagram of key components of the EVA.

Table 3

MOX sensors for the electronic volatile analyzer.

Sensor	Targeted gases	Commercial application	Manufacturer
GGS2330	CO, H2, ethanol	Wide range applications	Umwelt Sensor Technik, Germany
GGS1330	Hydrocarbon, combustible gases	Gas leak detection	Umwelt Sensor Technik, Germany
TGS2611	Methane	Gas leak detection	Figaro USA, Inc., USA
TGS2602	VOCs and odorous gases such as ammonia and H2S	Indoor air quality monitoring	Figaro USA, Inc., USA
TGS2600	H2, ethanol, air pollutants	Indoor air quality monitoring	Figaro USA, Inc., USA
TGS8100	H2, ethanol, air pollutants	Indoor air quality monitoring	Figaro USA, Inc., USA

Each MOX sensor was operated using an individualized multi-step, periodic voltage profile applied to the heating element, which resulted in stepwise modulation of the device temperature (Fig 2). Modulation of the temperature of MOX sensing elements is a well-known technique that can be used to enhance and tune the dependence of the sensing element resistance to the surrounding environment. The heater voltage profiles applied to the EVA sensors were optimized to maximize the sensitivity of each sensor with respect to a subset of target VOC, as well as to improve response orthogonality. Although the duration and amplitude of the individual waveforms was adjusted independently for each sensor, all waveforms were synchronized to a period of 80 s to simplify the handling and processing of the sensor array outputs. The resistance of each MOX sensing element was monitored at a fixed voltage and a rate of 10 Hz. During operation, a constant flow of 150 ± 10 sccm was established by means of a small pump, drawing air from the environment in a continuous fashion while also collecting the headspace of the sample of interest. The vapor-carrying flow was directed towards an enclosed chamber containing the six MOX gas sensors for the exposure and detection to take place.

Fig 2

Example of a temperature profile modulation for MOX sensors for IBM EVA™: (i) periodic waveform of heater voltage, expressed as a percentage of the maximum operating voltage recommended by the sensor manufacturer; (ii) corresponding variations in MOX sensor resistance under constant environment.

Chemicals and reagents

Normal Human Urine was purchased from UTAK Laboratories Inc., USA. The vendor obtained consent from the donors, who were healthy and drug-free. Two strains of E. coli were purchased from the American Type Culture Collection (ATCC). A nonpathogenic strain (ATCC 29425) is designated as K12, and the other is a uropathogenic strain (ATCC 700928) designated as CFT073. They are, henceforth, referred to as K12 and UPEC, respectively. Bacterial growth medium was prepared from Tryptic Soy Broth (TSB) powder (BD Bacto). For GC-MS, HPLC-grade ethanol (EtOH) and isopropyl alcohol (IPA) were obtained from Sigma Aldrich, USA.

Preparation of EVA samples and GC-MS reagents

After thawing, the urine bottle was vigorously shaken and aliquoted into multiple 50-ml centrifuge tubes. Upon usage, the tubes were centrifuged at 15,000 rpm for 15 minutes to concentrate any visible sediments, which were then removed by filtration. The resultant urine was named filtered normal urine or fNU.

The TSB medium was mixed and autoclaved using the standard method printed on its bottle. The powder form of the bacteria was resuspended in liquid TSB. Each suspension was then streaked on a Tryptic Soy Agar using standard aseptic techniques. The presence of many colonies after 24 hours of incubation at 37°C indicated bacterial viability. The handling of the UPEC was carried out by BSL-2 trained personnel under a BSL-2 cabinet. The agars with colonies were stored at 4–7°C and re-streaked every two weeks to keep the cell line fresh.

One colony of each E. coli strain was aseptically inoculated in separate tubes of 10 ml of fNU. The cultures were incubated for 24 hours before being filtered using the same procedure for fNU, except the centrifugation speed was 5000 rpm to avoid the destruction of the bacterial cells that can inadvertently introduce unwanted substances to the supernatant. Five milliliters of the supernatant were transferred to a septa-top vial and kept in the refrigerator at 4–7°C until analysis.

A stock solution for EtOH at a 1/100 dilution was prepared by adding 1 ml of pure, 200-proof ethanol in 9 ml of Millipore water in a tube labeled 1/10 EtOH. After vortexing, 1 ml of the 1/10 EtOH was added to 9 ml of Millipore water in a tube labeled 1/100 EtOH and vortexed. The 1/100 solution of IPA was prepared in a similar manner from a bottle of pure IPA. The equivalent concentration in g/ml and ppm for both solutions is 0.008 g/ml and 8 ppm.

EVA measurement and two-fold cross-validation ANN

Samples were prepared in Wheaton Septa-top Vials with a rubber-top cap that can be punctured with two needles to allow air intake. The first needle is connected to the EVA by a silicon tubing for headspace flux, and the second is connected to a HEPA carbon filter (0.22 μm) and left open to lab air for venting. Four samples were measured with EVA: an empty vial as lab air, 5 ml of normal fNU, 5 ml of K12, and 5 ml of UPEC. Each vial was measured sequentially at a sampling rate of 10 Hz. Additionally, the measurement sequence was randomized to minimize any history effects. Each sample was measured once in each cycle for ten minutes in a continuous flow and reiterated after all other samples have been measured. Between each vial measurement, the device was also flushed for five minutes to purge off residual VOCs.

With temperature oscillation resulting from the heater voltage waveform over 80 seconds, the six sensors give rise to a total of 120 features. Feature extraction was achieved through an amplitude-driven approach of extracting the mean area under the curve for the given response duration using GNU Octave v-5.1.0.0 (GUI), available through open source. Feature extraction every 80s constitutes one training example for the ANN model for a total of 7 examples per sample per cycle. Upon determining the training and testing sets by randomly splitting the overall dataset in half, both datasets are fed into a backpropagating ANN with two hidden layers of 24 and 9 nodes, respectively, which reports the training and testing results. The accuracy was calculated by dividing the total number of correct predictions over the total number of examples. Second-fold cross-validation was performed by switching the training and testing sets and repeating the analytics. The robustness of the ANN model was assessed with both accuracies.

Preparation of calibration standards

Calibration standards were prepared by spiking urine with certain amounts of EtOH and IPA. The final alcohol concentration was calculated using the following equation: where C is the concentration in ppm; X is the dilution factor, which is 1/100; D is the density of the alcohols, which is 0.8 g/ml for both IPA and EtOH; Valc is the volume in μl of the alcohols and Vurine the volume in μl of urine.

A preliminary investigation found that the EtOH in UPEC-inoculated urine was approximately 20 to 30 ppm using a non-validated headspace GC-MS analysis with random EtOH calibrators. Based on this result, a more defined range of concentrations for the EtOH calibrators were determined to be 10, 15, 20, 30, 50, 75, and 100 ppm. The final volume reflects the ultimate amount after transferring some amount to make lower dilutions (e.g., 12.0 ml of the 100-ppm concentration was used to make the 75-ppm concentration). Approximately 16 milliliters of each concentration were prepared so that they could be split into three batches for subsequent inter-day validation. The exact volume for each concentration was carefully calculated and tabulated in Table 4. After aliquoting 5 ml of each concentration into septa-top vials, each vial was spiked with 31.25 μl of IPA for a final concentration of 50 ppm IPA.

Table 4

Serial dilution of calibration standards.

Target Concentration (ppm)	Urine Volume (ml)	EtOH Volume (ml)	EtOH Source	Final Volume (ml)
100	47.4	0.6	1/100 Stock	16.0
75	4.0	12.0	100 ppm	16.0
50	20.0	20.0	100 ppm	16.2
30	11.6	17.4	50 ppm	15.7
20	9.6	6.4	50 ppm	16.0
15	8.0	8.0	30 ppm	16.0
10	10.6	5.3	30 ppm	15.9

Preparation of quality control standards

Three Quality Control (QC) standards were prepared at 25 ppm, 60 ppm, and 90 ppm and named QCL, QCM, and QCH, respectively, for low, medium, and high concentrations. Ideally, the QCL should be at the maximum of three times higher than the lower limit of quantitation (LLOQ), which is ten ppm. However, the QC level should not repeat any calibration concentration, so the 25 ppm was determined. They were prepared by directly spiking the urine with the 1/100 ethanol stock. Six QC vials were required, so 32 ml of each QC level was spiked as in Table 5.

Table 5

Preparation details for QC standards.

QC Name	Concentration (ppm)	EtOH Volume (ml)	Urine Volume (ml)	Final Volume (ml)
QCH	90	0.360	31.64	32
QCM	60	0.240	31.76	32
QCL	25	0.100	31.90	32
LLOQ	10	0.040	31.96	32

The 32-ml QCs were then dispensed into six vials at 5 ml per vial and spiked with 31.25 μl of IPA for a final concentration of 50 ppm IPA. The same QC preparation shall be repeated for each subsequent day of validation. For a full three-day validation, approximately 500 ml of filtered urine is required. The method was validated for its specificity, linearity, intra-day and inter-day inaccuracy and imprecision. The validation method and acceptance criteria were according to Bioanalytical Method Validation, published by the US Food and Drug Administration [31].

GC-MS parameters and conditions

The GC-MS used was an Agilent 6890 GC system equipped with HP 5973 Mass Selective Detector. Detection was done using a quadrupole detector with an electron impact source. The analytical run was performed on a DB ALC column with 30-m length, 0.32 internal diameter, and 1.80 film thickness. Inlet temperature was set at 100°C. The oven temperature was initially held for two minutes at 40°C then ramped for 25°C/min until 250°C and held for another two minutes, and the total running time was 26 minutes under the scanning mode. Under the selected ion monitoring (SIM) mode, inlet temperature was set at 100°C and the column temperature at 40°C without temperature ramping. The column flow was set at 6 mL/min with a total runtime of six minutes. The split mode was used with a split ratio of 25:1. Helium was used as a carrier gas with a flow rate of 0.9 mL/min. Isopropyl alcohol was used as an internal standard (ISTD). A Hamilton 1005SL 5-ml gas-tight syringe was used to withdraw and inject the headspace samples into the GC-MS manually.

Before the injection, the sample was placed in a heat block that had been heated to approximately 80°C and kept incubating for ten minutes to increase headspace VOCs (S1 Fig). A metal blockade was placed on top of the vial cap to prevent accidental touching of the syringe tip and the liquid. Five milliliters of the headspace sample was injected into GCMS under scanning mode as described previously to determine the compound of interest. After injection, the syringe was thoroughly cleaned with water and dried with an air gun.

Assessment of sample stability after EVA measurement

Sample evaporation becomes an issue during EVA measurement due to continuous suction of the headspace volume. It was observed that the sample volume got reduced over time. Therefore, the newly developed method was used to quantitate the stability of ethanol content in K12-inoculated urine during measurement.

Due to the destructive nature of the GC-MS assay, multiple spiked samples were prepared to be measured at different time points (S2 Fig). A stock of 25-ppm ethanol-spiked urine stock was prepared and split into nine vials. One vial was quantitated with the GC-MS to confirm the initial concentration. The remaining eight vials were divided into two groups: one group was measured by the EVA, and the other group was not (control). One vial from each group was quantitated by GC-MS after each EVA measurement cycle of ten minutes.

Results and discussion

Specificity

The mass over charge number ratios (m/z) 31 and 45 for ethanol and the m/z 45 and 43 for isopropyl alcohol were selected based on pre-scanning of ethanol and IPA. Blank urine sample, blank urine sample spiked with the internal standard, and blank urine sample spiked with 10 ppm EtOH and 50 ppm ISTD were prepared and injected into GC-MS. Two separate peaks were observed for EtOH and ISTD (Fig 3). In particular, ethanol was detected at 4.38 min and IPA at 5.45 min. No significant interfering peaks were found at the retention times at which ethanol and IPA appears. The signal to noise ratios for both drugs was greater than 10. The results demonstrate the adequacy of the method for the specificity of the compounds involved.

Fig 3

Chromatograms of (a) blank urine, (b) urine sample spiked with 50 ppm IPA, and (c) urine sample spiked with 10 ppm ethanol and 50 ppm IPA.

Linearity

A calibration curve was plotted for each validation day for the ratio of EtOH to ISTD response against the concentrations of ethanol (Fig 4). The linearity was observed for all calibration curves performed during method validation with all R2 values above 0.990 (Table 6).

Fig 4

A representative standard curve.

The y-axis plots the response ratio between the internal standard and the analyte.

Table 6

Standard curve equations and their coefficients of determination.

Validation assay	Linear equation	R2
Day 1	y = 0.0064x + 0.0096	0.9957
Day 2	y = 0.0065x + 0.0068	0.9916
Day 3	y = 0.0063x + 0.0134	0.9916

Within-assay reproducibility

The mean concentration from six replicates of each QC level was calculated, along with their standard deviation and coefficient of variation (Table 7). The mean inaccuracy was calculated by averaging the percent difference between each data point and the mean value. All inaccuracy values were within acceptable ranges according to US FDA guidelines (20% or less for LLOQ and 15% or less for other QCs).

Table 7

Within-assay coefficients of variation and mean inaccuracies.

Batch number	QC	Nominal conc. (ppm)	Mean (n = 6) (ppm)	SD (ppm)	CV (%)	Mean inaccuracy (%)
1	LLOQ	10	8.50	0.83	9.76	15.00
	QCL	25	24.24	1.31	5.40	4.9
	QCM	60	61.59	3.19	5.18	4.74
	QCH	90	98.01	4.77	4.87	9.01
2	LLOQ	10	8.70	1.07	12.30	15.62
	QCL	25	22.98	2.44	10.62	11.89
	QCM	60	55.68	7.47	13.42	12.92
	QCH	90	87.47	5.63	6.44	5.49
3	LLOQ	10	11.06	0.97	8.77	10.63
	QCL	25	25.07	3.30	13.16	8.80
	QCM	60	62.38	3.45	5.53	4.79
	QCH	90	89.87	4.63	5.15	4.25

Between-assay reproducibility

The between-assay repeatability was assessed by calculating the mean, the standard deviation, the CV, and the mean inaccuracy across all 18 samples from the three batches. All mean inaccuracy values were within acceptable ranges according to FDA guidelines (Table 8).

Table 8

Between-assay coefficients of variation and mean inaccuracies.

QC	Nominal conc. (ppm)	Mean (n = 18) (ppm)	SD (ppm)	CV (%)	Mean inaccuracy (%)
LLOQ	10	9.42	1.5	15.92	13.75
QCL	25	24.1	2.5	10.37	8.56
QCM	60	59.88	5.69	9.50	7.48
QCH	90	91.79	6.62	7.21	6.25

Quantitation of ethanol in E. coli-inoculated urine

As mentioned in the introduction, there are numerous studies on the e-nose application in detecting and differentiating the causative pathogenic species. Interestingly, most of these studies looked at the VOCs of bacteria in a nutrient agar or jelified urine instead of directly measuring the headspace of the urine sample. Presumably, the solidified form could minimize the evaporation of background components since the presence of water molecules and volatile nutrient broth components could have introduced variability to the sensor system. It is thus important to demonstrate the differentiation capability of an e-nose directly within the headspace of liquid samples for better representation as a point-of-care device. Aathithan et al. analyzed the direct urine samples but did not comment on sample quality, and the reported sensitivity and specificity were not impressive, given that the classification is either positive or negative infection without further strain identification. Here, we devised a benchtop model of UTI by infecting the normal urine samples because it is easy to access, establish, and control in terms of bacterial culture (K12 vs. UPEC) and ethanol concentration. Patient-derived samples could have extremely varied bacterial profiles as well as VOC profiles; that would be beyond the scope of this study. Using a benchtop model, we can decouple confounding factors to understand better the contribution of data quality in discriminating between species. Therefore, the GC-MS method was used to measure samples of K12-inoculated and UPEC-inoculated urine in tandem with an artificial neural network classification of odor data collected by the EVA.

The GC-MS quantitation resulted in 31.33 ppm of ethanol in UPEC-inoculated urine and 18.00 ppm in K12-inoculated urine. Besides the two abovementioned labels, the EVA also measured lab air and normal urine. The dataset was split in half for the training and testing set and evaluated with two-fold cross-validation. The classification resulted in 100% accuracy for both validations, which was determined by taking the ratio between the number of correctly predicted examples over the total number of examples. The VOC fingerprints are visualized by plotting the log of the electrical resistance across all features (Fig 5). The results thus highlighted the potential of EVA to distinguish bacterial strains directly on liquid samples.

Fig 5

Visualization of VOC fingerprints shows different patterns in electrical resistance across calculated sample features for each sensor.

These results demonstrated that UPEC and K12 E. coli produced distinguishable levels of ethanol, thus causing different VOC fingerprints. The four color-coded labels clearly show different patterns from one another when taking all feature responses as a whole. Without extensive scans, it is not known whether the two strains also have distinct levels of other VOCs. However, the difference in compound concentration enables the differentiation of these two strains of the same species. The GC-MS method allows researchers to directly study the VOC profiles in inoculated urine and aids in decision-making regarding establishing a better measurement protocol and choice of training data for e-nose application.

Sample instability and consideration for EVA measurement time

The concentration of ethanol-spiked urine was quantitated after every EVA cycle for four cycles. Adhering to the FDA acceptance range, a percent change of 15% or more was deemed significant. Fig 6 shows that the concentration of ethanol diminished significantly (more than 15%) as early as after the third cycle, which is equivalent to 30 minutes of measurement per vial.

Fig 6

Percent change in ethanol concentration after each EVA measurement cycle.

Next, a new classification was performed using only three labels of lab air, normal urine, and K12. Data were collected up to 15 cycles each day for two days. The data from the first three cycles were used as a training set to test every subsequent set of three cycles. The results in Table 9 indicate that accurate classification was possible up to nine cycles. Therefore, there are two thresholds for measurement cut-off: the first three cycles and the first nine cycles.

Table 9

Classification accuracies from training with the first three cycles.

Tested on:	Cycles 4–6	Cycles 7–9	Cycles 10–12	Cycles 13–15
Average accuracy:	95.70%	98.20%	55.50%	37.10%

Instability in biological samples, especially urine, has been investigated and shown that long storage time can reduce the emitted VOCs as measured by an e-nose [23]. Typically, the first few examples taken by e-nose are discarded during data processing to avoid variability due to sensor drifting and initial sample instability [32]. However, little is known about sample instability during e-nose measurement. Roine et al. suggested the measurement time can be reduced to 5 minutes based on their classification results [6]. However, this suggestion can only be made after the fact without a priori quantitative basis. Also, a short measurement time gives fewer examples for training the e-nose, a trade-off that needs to be carefully considered. Here, we determined the appropriate measurement time then tested whether using the data within the specified cut-off would yield a better classification than including data after the cut-off.

Between-day classifications were used to evaluate the two cut-offs, and the results are plotted in Fig 7. Training with the first 30 minutes (three cycles) of data gives better accuracies of 97.0% and 73.7% compared to training with the first 90 minutes of data.

Fig 7

Cross-validation accuracies between days for 90-min and 30-min measurements.

Thus, sample instability was induced after 30 minutes of measurement with an e-nose. Specifically, the ethanol concentration was reduced by more than 15%, which is significant according to the FDA guidelines. It was further demonstrated that the ANN model gave superior classification accuracies when using only data from the first 30 minutes of measurement. The result is an example of machine learning basics: data quality is more important than data quantity. The inclusion of data from unstable samples is detrimental to the classification model. The GC-MS method was thus proved useful for determining the measurement cut-off time.

Here, each measurement cycle lasts for ten minutes, which agrees with other studies in the literature [6, 13]. However, there is no standardized protocol for how long this process should be. In a more recent study by Capelli et al., the urine headspace was flown into the sensor chamber for 50 minutes [33]. It is thus possible to rerun sample collection to increase the number of training examples for the model as long as the cut-off time can be determined, which could improve the efficiency of this assay.

Since e-noses typically output electrical resistances, which result from the collective interaction of the whole VOC profile with the sensors, they lack the credibility to aid in decision-making regarding quality control of the data. Furthermore, we recognize that the VOC profiles in real patient-derived UTI samples are likely more complicated. For example, E. coli is not the only species that produce ethanol as a by-product, but other UTI-causing strains such as Klebsiella aerogenes, albeit less prevalent, also ferment lactose and release ethanol. It would be beneficial to strengthen the sensitivity and specificity of the sensors toward ethanol. However, the EVA will be designed to distinguish various bacterial strains, many of which may produce compounds other than ethanol. Thus, we must also include a variety of sensors to account for this dynamicity. These sensors are modular and can be added, removed, or replaced. Therefore, we could optimize a sensor array that targets the most prevalent strains associated with UTI.

There are a number of limitations to this study that should be acknowledged. The most notable one is that we tested the EVA on inoculated urine in developing our assay. While providing consistency in sample preparation in a proof-of-concept setting, we cannot project the results for clinical samples in which urine can vastly vary in texture and contents, causing a lot of noise for the neural network. It is important to analyze the differential concentration of VOCs in healthy vs. UTI to screen for appropriate sensors that respond to the top compounds with the most difference. A second limitation is the drifting of sensors. We did not specifically assess the age of the sensors used in our experiment. Bax et al. showed that classification using one-year-old sensors was much worse than using new ones, and they proposed a correction model that significantly improved the performance from 55% to 80% [34]. Regarding drift, baseline shift among analyses performed on different days could also be seen as a shortcoming of the study, as ambient temperature or humidity could influence baseline sensor readings. We did not observe a significant shift in baseline over the experimental duration. However, Bax et al. also described a pre-treatment procedure to compensate for baseline drift by using Standard Normal Variate. The same procedure can be adapted to improve our model. A third limitation is that the testing time of 10 minutes for each sampling cycle could have been a little too long. As aforementioned, a shorter test time has been proposed [6]. We could further reduce the length of a single measurement from 80 seconds.

Future studies should carefully evaluate the degree of sample degradation. While sample instability is inevitable, it is efficient to determine the extent of this degradation to maximize data quantity without a trade-off for data quality. The same headspace analysis should be considered in every future experimental design to determine the ground truth, which is the concentration of analytes, before proceeding to use the collected data. By characterizing the sample stability during measurement, one can safely reduce variability in the data due to the very act of measurement itself while being able to incorporate background compounds as part of the VOC profile.

Conclusion

In this paper, we demonstrated the validation and quantitation data of ethanol in the headspace of urine samples inoculated with E. coli. The method validation fulfilled all the criteria as outlined in the Bioanalytical Method Validation protocol published by the US FDA. The method was successfully applied on ethanol measurement in samples of K12 and UPEC-inoculated urine as an initial step towards improving the outcome of VOC measurement by electronic nose technology through a validated headspace GC-MS method. The main interest of using this method was to characterize E. coli-inoculated urine samples that are prepared for training an electronic nose to detect urinary tract infection and differentiate between different causative agents. By using the new method, it was shown that different strains of E. coli could produce different levels of ethanol concentration, making it possible to differentiate between them based on distinct VOC fingerprints. The quantitation also revealed that e-nose measurement could affect sample stability over time. Therefore, the cut-off time for future measurement should be wisely determined to avoid the collection of unusable data due to a reduction in compound concentration. The GC-MS will continue to be a useful tool to support technology development, in characterizing samples to build a useful training library for UTI detection directly through liquid samples without extra preparation steps; thus enabling next generation real-time and point-of-care diagnosis of UTI.

Headspace sample being withdrawn from urine vial heated in a heat block.

The real temperature was measured in an adjacent water vial by a digital thermometer. Inset: Urine sample in a septa-top vial for headspace analysis.

(TIF)

Click here for additional data file.

A schematic of urine sample stability test with GC-MS.

(TIF)

Click here for additional data file.

Sample set up with the EVA.

(TIF)

Click here for additional data file.

6 Dec 2021

A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.

A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.

An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols.

We look forward to receiving your revised manuscript.

Kind regards,

A. M. Abd El-Aty

Academic Editor

PLOS ONE

Journal Requirements:

When submitting your revision, we need you to address these additional requirements.

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at

https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf.

2. Please include in your Data availability statement details of how other researchers may access the prototype information and any equipment or data used in this study which is not public. Please include information on how another researcher would enter into a a Joint Study Agreement with IBM Research, plus a non-author IBM contact for facilitating that process. Please also clarify whether the authors had any special privileges in accessing proprietary equipment or data."

3. Please complete the Financial disclosure statement and Competing interests statement, including all financial and materials support, as well as any author commercial affiliations."

4. Please include a caption for figure 8.

5. Please ensure that you refer to Figure 7 in your text as, if accepted, production will need this reference to link the reader to the figure

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: Partly

Reviewer #3: No

Reviewer #4: Yes

**********

2. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: No

Reviewer #3: Yes

Reviewer #4: Yes

**********

3. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: No

Reviewer #2: No

Reviewer #3: Yes

Reviewer #4: Yes

**********

4. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: No

Reviewer #3: Yes

Reviewer #4: Yes

**********

5. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: “Quantitation of ethanol in UTI assay for volatile organic compound detection by

electronic nose using the validated headspace GC-MS method” is an excellent study where the authors used inoculated urine samples as a proof-of-principle to show the ability of the EVA IBM e-nose to distinguish between two E. coli species. Interestingly, the authors explored the relationship between volatile organic compound (VOC) - ethanol - and e-nose result accuracy. The neural networks of e-noses do not offer an insight into the composition of the headspace. In contrast, GC-MS can shed light on the VOCs present – using this tool the authors show that just testing the sample with the e-nose degrades the quality of the sample over time. Thus, one of the most important takeaways from this study is that the quality of the data is more important than the quantity of data. Furthermore, this study advances the field, paving the way to faster and cheaper UTI diagnosis using e-noses by showing the abilities of the EVA to accurately discriminate between samples. Simultaneously, it builds up a GC-MS method to study headspace VOC with the view of gathering data for better choice and calibration of the sensors in the e-nose.

This study follows up on a conference paper in 2019: “Rapid Strain Differentiation of E. coli-inoculated Urine Using Olfactory-based Smart Sensors”. It was interesting to see the progress done since then. The manuscript submitted by the authors is the only use of EVA IBM e-nose on urine samples I could find in the literature.

Minor

There are some points that need to be addressed:

1. The calculations in table 4 are correct based on the formula provided in line 236 (the whole paragraph, line 234-239, is very useful – well done for including!). However, the final volume column is confusing as sometimes the urine volume + EtOH volume does not add up to the final volume. For example, target concentration 100 ppm 47.4 (Urine volume) + 0.6 (EtOH volume) is 48 ml not 16 ml (Final volume written in the table). Similarly, this happens again for target concentration 50 ppm and 30 ppm.

The other target concentrations do add up to what is in the table. I think this needs to be clarified.

2. Table 5 (line 266) appears to have an error: the EtOH volume appears to be wrong – when used for calculation it does not give the target concentration or the final volume (EtOH volume 0.008 ml + urine volume 31.96 volume = 31.968 ml total volume; not 32 ml as written in the table). The correct concentration in ppm, as well as the final volume, works out with 0.04 ml of EtOH.

3. Line 94 “Table 1 is limited to those used in urinary pathogen detection” and line 357 states that e-nose studies previously looked at nutrient agar or gelified urine instead of direct urine measurements.

There have been previous e-nose studies that use urine directly or looked at urinary pathogen detection such as: Aathithan, S.; Plant, J.C.; Chaudry, A.N.; French, G.L. Diagnosis of Bacteriuria by Detection of Volatile Organic Compounds in Urine Using an Automated Headspace Analyzer with Multiple Conducting Polymer Sensors. J. Clin. Microbiol. 2001, 39, 2590–2593.

A review that summarises the literature is available: Dospinescu, V.-M., A. Tiele & J. A. Covington (2020) Sniffing out urinary tract infection—Diagnosis based on volatile organic compounds and smell profile. Biosensors, 10, 83.

Table 3 in the aforementioned review encapsulates the studies published in the field and provides an ampler overview of the experiments previously conducted on VOCs and infected urine/pathogens involved in UTI.

4. Presentation/Grammar:

The paragraph between lines 240-246 is repeated (lines 247-253).

Line 284 - “temperature was set at 100oC” should be 100°C.

Line 71 – “urines” should be urine.

Once again, I believe this study advances the field and shows that the IBM EVA e-nose has the potential to be used as a diagnostic tool for UTI with enough developments. In addition, it also provides a framework that could be used for other VOCs (not just ethanol) for optimising e-nose use based on GC-MS data. My recommendation is that the study should be published once the points above are addressed.

Other comments:

1- The text is sufficiently detailed to understand the work whilst also being succinct. It reads easily and keeps the reader engaged. The problem is framed in the introduction and the aim is identified in lines 105-107.

2- E. Coli ferments lactose (or arabinose) and as a by-product it produces ethanol. However, this is not specific to E. coli, another agent that causes UTI (although not as common) also produces ethanol: Klebsiella aerogenes.

3- Data availability:

The authors state “Yes - all data are fully available without restriction” but in the next section “Data access are available upon request” together with two links (“GC-MS data are available at

https://app.box.com/folder/125090388794 and Electronic nose EVA data at

https://app.box.com/folder/113625571232”).

However, even after signing up on the website, accessing the link gives the following error “Oops! We can't seem to find the page you're looking for.” - Considering how "The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction" this app.box issue should be addressed.

Reviewer #2: The paper reports on the use of an electronic nose for measuring ethanol concentration in urine samples. It is really about using validation methods on artificial urine samples and I think the title is a little miss leading as no samples from UTI sufferers are included. A large percentage of the paper is given over to GCMS and the application of a standard to these measurements. The authors should be clear that this is really the purpose of the paper and change the focus on the paper. I then feel it would be more interesting to the community. The introduction needs a fair amount of work and some of the statistics need checking. Also, some of the purposes of the tests need to be explained in context to UTI. For example, a 30 minute electronic nose measurement is a very long time. I would like to see more common electronic nose analysis – so a simple PCA of the features would be useful. In terms of sample re-runs, this is not something that you would undertake with an electronic nose and there are a number of papers that discuss sample testing optimisation (which are not referenced here). The paper needs to be re-structured and re-written with a different focus.

Some specific comments are below.

Line 50: It is a little odd to start a urine based study mentioning breath. Maybe this could just list the biological sources?

Line 54: Poor English. Also no details are provided about the study. Was it an infection or kidney failure or?

Line 59: Was this of just the bacteria or an infected human sample?

Line 62: Don’t understand what you are trying to say.

Line 65: I disagree with this statement. A 10 cent dipstick test will inform you if you have UTI. More important would be to identify the bacterial species or identify UTI in those where there is a high false positive rate.

Line 69: I disagree with this statement. How can static headspace analysis be better that a trap pre-concentration step? Or do you mean something else? Also, I am not convinced this paragraph adds anything. I don’t think you need to prove that headspace analysis is not a valid approach.

Line 83: You should give a better explanation (or more classical explanation) of what an electronic nose is and then remove table 1. I would give examples in the introduction of previous UTI student and then compare your result with the literature in a table at the end of the work. I am unsure what the relevance of a prostate cancer study, but you could mention earlier the relevance of cancer urine studies (bladder, prostate, colorectal etc.).

I don’t think table 2 is relevant. I would like to see some focus on previous literature showing that ethanol is important (which isn’t discussed) and how it is modulated in the presence of disease. Also, some comment on the biological pathway that creates ethanol – from host response or from the bacteria itself. This should be included in the introduction.

Line 144: You are working on detecting ethanol, but many of the sensors are not targeting ethanol, why is this? Also, it is an odd choice of sensors. Was there a reason for this combination from different manufacturers? Was any optimisation of the array undertaken?

Line 158: Please provide technical details of how the unit was driven and measured. For example, you provide heater voltage as a %, not as a V. Where all the sensors operated with the same temperature pulses or were they different?

Line 163: How was this optimisation undertaken?

Line 168: How was resistance measured with a fixed voltage? What was the internal volume of the chamber? Was there a background reading before the measurement or were you just using the temperature modulation to give you an non-sensing resistance? Or was something else done?

Line 216: Please provide details of the vials and how they were modified/used to allow an air intake. What was the tubing used?

Line 223: Not clear how you extracted features from the raw data.

Line 227: What software/program/model was used to create the BNN?

Line 229: Were the samples from the same sample excluded from the training set for when they were being used as a test set? Otherwise, you are training and testing on the same samples. It is not clear how you are doing the cross-validation.

Line 240: How was this done?

Line 271: Was this human urine or artificial?

Line 323: Figure 3 is really difficult to read and there doesn’t appear to be any axes labels on the figure (though it might be the poor quality of the image). I wonder if the PDF process has caused this?

Line 358: I am pretty sure there are some UTI studies using direct analysis. The authors should comment on these papers as well.

Line 397: Why is this important? Would not the sample be tested and then disposed of? If a second reading was needed, they would just take some more out of the sample container – or just get more urine from patient. The reason for doing this needs to be explained. You are providing evidence that you should just test once.

Line 415: Why would you measure for 30 minutes with an electronic nose? For what purpose?

Line 427: There have been a number of studies looking at urine stability with electronic noses – which I noticed are not referenced. Also, the result found here is well known in the electronic nose community.

Line 437: This is really important – and much more that the focus on electronic noses. I would like this to have been in the introduction.

Line 445: Would this not be dependent on the level of infection in real life?

Reviewer #3: The capabilities of quantifying individual VOCs is not normally an important requirement for identification of diseases or pathogens responsible for causing diseases when using electronic-nose devices. The most important information to validate is the identities of VOCs making up the E-nose smellprint signatures and thus VOC profiles, not concentration of VOCs which is normally only needed in metabolomic studies to determine effects of pathogens on metabolic pathways of the host. Thus, quantification does not add significantly to the capabilities for testing the efficacy of a new experimental e-nose device. The ultimate objective was to develop the capabilities of the e-nose for detecting UTI caused by different microbes. The methods developed here do not contribute to that objective and the data obtained is normally part of a pilot study for methods development and not published as a stand alone unit without e-nose data on UTI samples from different types of microbial causes with adequate controls. Quantification of a single possible VOC in a UTI sample headspace provides very little information towards development of an e-nose library database containing specific complex mixtures of VOCs that affect the output signatures of the e-nose sensor array. The sensor array responds to all of the VOCs present in the headspace, not just a single VOC such as ethanol.

All of the figures (Fig. 1-7) are of very low quality resolution and do not provide the data necessary to support the efficacy of a new experimental e-nose device for UTI diagnostics (based on quantitation of EtOH alone). Development of methods for quantitation of EtOH along with a standard curve alone do not contribute significantly towards development of e-nose methods useful for UTI diagnostics and thus the Conclusions are not supported by the objectives of the study or the data obtained towards this purpose.

Reviewer #4: The paper describes the use of an electronic nose to distinguish the pathogens causing urinary tract infection, and proposes an new experimental method to improve sample stability during e-nose measurements. The problem of sample stability during e-nose measurment should be better claimed in the introduction, because it is the focus of the described sperimentation.

The methods involved to prepare EVA samples and GC-MS calibrants are described properly. However, the authors should clearly state the reasons leading to the use of normal urine samples to be inoculated rather than real infected samples. I think that authors should also better describe the approach involved for training the e-nose. Specifically, they should describe the scheme involved for presenting samples to the e-nose and validating the classification performance. The paragraph (lines 223-231) is not clear. The value of the sensor resistance at different times was used as features?

All figures of the paper have a very poor quality. They should be revised. In some cases, the text is not readable.

Moreover, the authors should add reference supporting their statments throughout the paper.

**********

6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

Reviewer #3: No

Reviewer #4: Yes: Carmen Bax

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step.

Submitted filename: PONE-D-21-22287 (1).pdf

Click here for additional data file.

28 Apr 2022

RESPONSE TO REVIEWERS (PONE-D-21-22287)

We appreciate the reviewers’ thoughtful comments and insight. We are pleased to receive many positive comments. However, we acknowledge several concerns raised by the reviewers. We have addressed these shortcomings in the revised manuscript. Response to reviewers is detailed as follows. We look forward to hearing from you regarding our revision. We would be glad to respond to any further questions and comments.

Reviewer 1:

1. "Quantitation of ethanol in UTI assay for volatile organic compound detection by electronic nose using the validated headspace GC-MS method" is an excellent study where the authors used inoculated urine samples as a proof-of-principle to show the ability of the EVA IBM e-nose to distinguish between two E. coli species. Interestingly, the authors explored the relationship between volatile organic compound (VOC) - ethanol - and e-nose result accuracy. The neural networks of e-noses do not offer an insight into the composition of the headspace. In contrast, GC-MS can shed light on the VOCs present – using this tool the authors show that just testing the sample with the e-nose degrades the quality of the sample over time. Thus, one of the most important takeaways from this study is that the quality of the data is more important than the quantity of data. Furthermore, this study advances the field, paving the way to faster and cheaper UTI diagnosis using e-noses by showing the abilities of the EVA to accurately discriminate between samples. Simultaneously, it builds up a GC-MS method to study headspace VOC with the view of gathering data for better choice and calibration of the sensors in the e-nose. This study follows up on a conference paper in 2019: "Rapid Strain Differentiation of E. coli-inoculated Urine Using Olfactory-based Smart Sensors". It was interesting to see the progress done since then. The manuscript submitted by the authors is the only use of EVA IBM e-nose on urine samples I could find in the literature.

We appreciate the positive comments from Reviewer 1 on the merit of this study and the promising perspectives of our EVA e-nose. Indeed, the key takeaway in this article is that the quality of the data could be a function of the data collection method and thus should be carefully controlled.

Minor comments

2. The calculations in table 4 are correct based on the formula provided in line 236 (the whole paragraph, line 234-239, is very useful – well done for including!). However, the final volume column is confusing as sometimes the urine volume + EtOH volume does not add up to the final volume. For example, target concentration 100 ppm 47.4 (Urine volume) + 0.6 (EtOH volume) is 48 ml not 16 ml (Final volume written in the table). Similarly, this happens again for target concentration 50 ppm and 30 ppm. The other target concentrations do add up to what is in the table. I think this needs to be clarified.

The final volume column reflects the ultimate volume after transferring some amount to make lower dilutions, not the initially generated amount. For example, at 100 ppm, 48 ml is initially generated, but a total of 32 ml (3rd column) is used to make the target 75-ppm and 50-ppm concentrations, hence the remaining 16 ml. We added a sentence to clarify this point.

3. Table 5 (line 266) appears to have an error: the EtOH volume appears to be wrong – when used for calculation it does not give the target concentration or the final volume (EtOH volume 0.008 ml + urine volume 31.96 volume = 31.968 ml total volume; not 32 ml as written in the table). The correct concentration in ppm, as well as the final volume, works out with 0.04 ml of EtOH.

We thank Reviewer 1 for pointing out this error. It has been fixed.

4. Line 94 "Table 1 is limited to those used in urinary pathogen detection" and line 357 states that e-nose studies previously looked at nutrient agar or gelified urine instead of direct urine measurements. There have been previous e-nose studies that use urine directly or looked at urinary pathogen detection such as: Aathithan, S.; Plant, J.C.; Chaudry, A.N.; French, G.L. Diagnosis of Bacteriuria by Detection of Volatile Organic Compounds in Urine Using an Automated Headspace Analyzer with Multiple Conducting Polymer Sensors. J. Clin. Microbiol. 2001, 39, 2590–2593.

We appreciate the reference suggested by Reviewer 1. Aathithan et al. using artificial urine inoculated with urinary pathogens and clinical urine samples is a great approach, making it a helpful reference. However, they used PCA with only two labels (infection positive or negative) which is not as powerful as the artificial neural network that we used. This point was reflected in the revised manuscript (lines 97-103).

A review that summarises the literature is available: Dospinescu, V.-M., A. Tiele & J. A. Covington (2020) Sniffing out urinary tract infection—Diagnosis based on volatile organic compounds and smell profile. Biosensors, 10, 83. Table 3 in the aforementioned review encapsulates the studies published in the field and provides an ampler overview of the experiments previously conducted on VOCs and infected urine/pathogens involved in UTI.

We are thankful for the very detailed table summarizing past studies on VOCs. We reflect this review in the introduction to highlight the potential of e-nose research (lines 65)

5. Presentation/Grammar:

The paragraph between lines 240-246 is repeated (lines 247-253).

Line 284 - "temperature was set at 100oC" should be 100°C.

Line 71 – "urines" should be urine.

We have fixed these mistakes.

Other comments:

6- E. Coli ferments lactose (or arabinose) and as a by-product it produces ethanol. However, this is not specific to E. coli, another agent that causes UTI (although not as common) also produces ethanol: Klebsiella aerogenes.

We appreciate this useful comment. Indeed, future studies can look more into the classification between K. aerogenes and E. coli and expand the profiles of VOCs measured by GC-MS to better design an e-nose capable of distinguishing them. We reflected this point in the Result and Discussion section (line 456)

7- Data availability:

The authors state "Yes - all data are fully available without restriction" but in the next section "Data access are available upon request" together with two links ("GC-MS data are available at

https://app.box.com/folder/125090388794 and Electronic nose EVA data at

https://app.box.com/folder/113625571232").

However, even after signing up on the website, accessing the link gives the following error "Oops! We can't seem to find the page you're looking for." - Considering how "The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction" this app.box issue should be addressed.

We have uploaded the data files to a public repository (Kaggle.com):

- GCMS data – raw files used to develop validation table - https://www.kaggle.com/datasets/aminatadebiyi/beva-plos-one

- IBM EVA data and features for Artificial Neural Network - https://www.kaggle.com/datasets/aminatadebiyi/beva-plos-one-ann

Reviewer 2

1. The paper reports on the use of an electronic nose for measuring ethanol concentration in urine samples. It is really about using validation methods on artificial urine samples and I think the title is a little miss leading as no samples from UTI sufferers are included.

We appreciate the valuable comment. However, we used normal human urine inoculated with E. coli to create a simple model of UTI, hence “UTI assay” in the title. The purpose of the paper is not to measure ethanol concentration but to show how sample quality can be affected by the very act of e-nose measurement and a proof of concept for distinguishing different strains of E. coli in inoculated urine samples.

2. A large percentage of the paper is given over to GCMS and the application of a standard to these measurements. The authors should be clear that this is really the purpose of the paper and change the focus on the paper. I then feel it would be more interesting to the community. The introduction needs a fair amount of work and some of the statistics need checking.

We clarify that the purpose of this study is to demonstrate the e-nose ability to distinguish K12 and UPEC E. coli in inoculated urine and that the classification could be further improved if sample quality can be controlled by the quantification of a characteristic VOC (ethanol). The GC-MS method development, therefore, is secondary to the main purpose of showing a successful classification.

We agree that the GC-MS method development takes a large portion of the paper, which is necessary because a specific method for the quantification of headspace ethanol in urine is uncommon. We switched a paragraph (lines 141) and revised the text regarding the purpose of the study to clarify the focus of the paper.

3. I would like to see more common electronic nose analysis – so a simple PCA of the features would be useful.

The reviewer is encouraged to check our previous report on the PCA and other classifiers. Adebiyi, A. et al. Sensors & Transducers Using Olfactory-based Smart Sensors. 238, 94–98 (2019). In this paper, we focused on demonstrating as a proof-of-concept the capability of an ANN model instead of comparing between multiple model, but we found the combination of our Feature Tuning approach with the ANN to maintain the robustness of our model over a longer period, probably due to sensor drift. Details of our feature tuning approach can be found here https://patents.google.com/patent/US20210172919A1/en.

4. Also, some of the purposes of the tests need to be explained in context to UTI. For example, a 30-minute electronic nose measurement is a very long time. In terms of sample re-runs, this is not something that you would undertake with an electronic nose and there are a number of papers that discuss sample testing optimisation (which are not referenced here).

We understand the concern. To clarify, each measurement lasts 10 minutes, yielding a total of seven training/testing examples (80 s per example) for the artificial neural network model. This time length agrees well with previous studies (e.g., Asimakopoulos et al. 2014, and Roine et al. 2014). A recent study even flowed the headspace into their eNose for 50 minutes (Capelli, L., Bax, C., Grizzi, F. et al. Optimization of training and measurement protocol for eNose analysis of urine headspace aimed at prostate cancer diagnosis. Sci Rep 11, 20898 (2021)). We updated the manuscript to reflect these references (line 247).

The 10-min measurement cycle was repeated several times to determine the optimal cut-off. By doing so, we achieve two goals. First, there was more frequent flushing in between to purge the sensor chamber. Second, we can maximize the number of training examples instead of discarding the sample after only 10 minutes of measurement. We show that we can include data up to 30 minutes (3 measurement cycles) before the sample degraded beyond the acceptable 15% fluctuation.

Specific comments

Line 50: It is a little odd to start a urine based study mentioning breath. Maybe this could just list the biological sources? Line 54: Poor English. Also no details are provided about the study. Was it an infection or kidney failure or?

We revised these lines to clear up unnecessary information.

Line 59: Was this of just the bacteria or an infected human sample?

It is of the bacteria themselves, according to a review by Bos et al., as mentioned in the sentence following that line.

Line 62: Don’t understand what you are trying to say.

We revised this sentence to clarify that even though multiple species can commonly produce a VOC, the headspace concentration could be different, thus providing the basis for bacterial discrimination.

Line 65: I disagree with this statement. A 10 cent dipstick test will inform you if you have UTI. More important would be to identify the bacterial species or identify UTI in those where there is a high false positive rate.

We agree with the reviewer that a high false-positive rate could contribute to the overuse of broad-spectrum antibiotics, in which case identification of bacterial species would be critical. However, the gold standard for UTI detection is bacterial culture which takes time. In many cases, patients can be given broad-spectrum antibiotics as a precautionary measure. A rapid test can thus lower this likelihood. A dipstick is also a rapid test, but it does not have specific identification like the e-nose. We reflected this point in the revised version.

Line 69: I disagree with this statement. How can static headspace analysis be better that a trap pre-concentration step? Or do you mean something else? Also, I am not convinced this paragraph adds anything. I don’t think you need to prove that headspace analysis is not a valid approach.

We are not proving that headspace analysis is not a valid approach or that it is better than a trap analysis. In this paragraph, we reason our choice for the static headspace method as it is relatively simple, has similar sensitivity to trap analysis or dynamic purge, and avoids time- and cost-consuming steps.

Line 83: You should give a better explanation (or more classical explanation) of what an electronic nose is and then remove table 1. I would give examples in the introduction of previous UTI student and then compare your result with the literature in a table at the end of the work. I am unsure what the relevance of a prostate cancer study, but you could mention earlier the relevance of cancer urine studies (bladder, prostate, colorectal etc.).

We appreciate the thoughtful suggestion. However, we think Table 1 is necessary to highlight sensor type (MOX) and algorithm (ANN) choice in our study. Since our scope is to highlight the GC-MS validation of our approach, we are happy to take out the table if it is confusing.

I don’t think table 2 is relevant. I would like to see some focus on previous literature showing that ethanol is important (which isn’t discussed) and how it is modulated in the presence of disease. Also, some comment on the biological pathway that creates ethanol – from host response or from the bacteria itself. This should be included in the introduction.

We added a reference for a review paper in the introduction (Dospinescu et al. 2019). This paper excellently reviewed the VOC profiles on relevant UTI-causing species. We commented on the ethanol production by E. coli in the introduction based on this paper.

Line 144: You are working on detecting ethanol, but many of the sensors are not targeting ethanol, why is this? Also, it is an odd choice of sensors. Was there a reason for this combination from different manufacturers? Was any optimisation of the array undertaken?

The IBM Electronic Volatile Analyzer™ is designed to be modular and is comprised of a variety of diverse sensors to respond to a variety of analytes, therefore to achieve orthogonality of sensors, we included the selection described. This specific combination of sensors was selected using an analytic-driven approached outlined in our patent “Sensor tuning - sensor specific selection for IoT-electronic nose application using gradient boosting decision trees.” https://patents.google.com/patent/US20210172918A1/en This combination of sensors and features provided the most robust classification for this biological application, which is why it was selected.

Line 158: Please provide technical details of how the unit was driven and measured. For example, you provide heater voltage as a %, not as a V. Where all the sensors operated with the same temperature pulses or were they different?

The design of the platform is modular: each gas sensor is mounted on its own sensor module, a Printed Circuit Board (PCB) of common design carrying an integrated microcontroller with built-in heating components, and the circuitry needed to operate the sensor. The sensor modules communicate via I2C protocol with a central hub (BeagleBone Black), a single-board computer that orchestrates the sensor modules and processes the multisensorial output.

Each MOX sensor was operated using periodic waveform of heater voltage, expressed as a percentage of the maximum operating voltage Vmax recommended by each sensor manufacturer. This resulted in a stepwise modulation of the device temperature (Figs. 2b(i)&(ii)), a well-known technique that can be used to enhance and tune the dependence of the sensing element resistance to the surrounding environment.

Although the duration and amplitude of the individual waveforms was adjusted independently for each sensor, all waveforms were synchronized to a period of 80 s to simplify the handling and processing of the sensor array outputs.

Line 163: How was this optimisation undertaken?

More details on our optimization techniques can be found at our patent filing, “Adaptive sensor temperature control for fast recovery,” highlighting our temperature control optimization method to maximize sensor response to our target analyte. (https://patents.google.com/patent/US20210063372A1/en)

Line 168: How was resistance measured with a fixed voltage? What was the internal volume of the chamber? Was there a background reading before the measurement or were you just using the temperature modulation to give you an non-sensing resistance? Or was something else done?

The data was collected at each given voltage at 10Hz, but the full sensor response was characterized as the time period of the operational voltage sequence (80s), that consisted of multiple voltage steps, giving us the response and recovery pattern of the sensor’s exposure to the sample analyte.

Rather than the use of background readings to minimize sensor drift, we performed system stabilization for an hour before each experiment. This involved exposing the sensors for an hour to lab air in ambient conditions. This stabilization method in combination with our choice of feed-forward neural network, captured the variation in our readings to maintain a robust system for the given measurements.

The interal volume of the chamber was about ~700ml.

Line 216: Please provide details of the vials and how they were modified/used to allow an air intake. What was the tubing used?

The vials used are Wheaton Septa-cap Vials. They have a rubber top cap that can be punctured with a needle. The tubing used was made from silicon. We provided an additional supplementary figure (S3) to show our set up.

Line 223: Not clear how you extracted features from the raw data.

Our feature extraction was achieved through an amplitude-driven approach of extracting the mean area under the curve for the given response duration. To focus the scope of the paper on the GC-MS validation, we did not go into much detail on this approach, however more details on this approach and the reasoning for its application can be found in our patent-filing, “Feature tuning – application dependent feature type selection for improved classification accuracy.” https://patents.google.com/patent/US20210172919A1/en

Line 227: What software/program/model was used to create the BNN?

GNU Octave v-5.1.0.0 (GUI), available through open source.

Line 229: Were the samples from the same sample excluded from the training set for when they were being used as a test set? Otherwise, you are training and testing on the same samples. It is not clear how you are doing the cross-validation.

We are thankful for pointing out the missing part. The dataset was split in half for the training and testing set (line 390), but we did not mention it earlier. We updated line 242 to reflect this part.

Line 240: How was this done?

We clarify (line 257) that the preliminary investigation was done with GC-MS, and the concentration was roughly estimated with a non-validated method using a non-validated headspace GC-MS analysis with random EtOH calibrators.

Line 271: Was this human urine or artificial?

Human urine was used as mentioned in line 194.

Line 323: Figure 3 is really difficult to read and there doesn’t appear to be any axes labels on the figure (though it might be the poor quality of the image). I wonder if the PDF process has caused this?

We believe the PDF process degraded the quality of our images, but we have updated the figure with a higher resolution image. We hope the quality is maintained after PDF processing.

Line 358: I am pretty sure there are some UTI studies using direct analysis. The authors should comment on these papers as well.

This is a useful suggestion. We included a new reference in the introduction. Aathithan et al. direct urine analysis but they use a simple PCA to detect either a positive or negative infection.

Line 397: Why is this important? Would not the sample be tested and then disposed of? If a second reading was needed, they would just take some more out of the sample container – or just get more urine from patient. The reason for doing this needs to be explained. You are providing evidence that you should just test once. Line 415: Why would you measure for 30 minutes with an electronic nose? For what purpose?

We appreciate the valuable suggestion. We hope that we have clarified the confusion on measurement time in response #4 above. Briefly, a longer measurement time can expand the number of examples collected as long as sample quality is maintained. Here, we collected data for training or testing the model three times (10 minutes each). We showed that beyond three times, the samples started to degrade, and these data should not be included. This way, we can maximize the inclusion of useful data without wasting samples (higher efficiency).

Line 427: There have been a number of studies looking at urine stability with electronic noses – which I noticed are not referenced. Also, the result found here is well known in the electronic nose community.

We revised our manuscript (lines 461) to include a paragraph discussing sample instability in general that is due to storage time and initial measurement with two new references (Ref. 23 & 24). We also referenced Roine et al. (Ref. 6) for their speculation that a short measurement time already allows for a good classification. However, we reason that this can only be made after the fact by classification results and that number of examples is a trade-off for shortening measurement time. Here, we devised a method to determine the optimal timepoint within which data can still be taken instead of a retroactive determination. We hope the urine stability has been well discussed and referenced accordingly. However, please let us know if the discussion needs further improvement and more references are needed. Although the results found here is well known in the electronic nose community, but the use of validated Headspace GCMS method to establish and optimise electronic nose measurement can be considered novel and the results from this study shows the usefulness of the headspace GCMS method in optimisation and quality control of ENose measurement.

Line 437: This is really important – and much more that the focus on electronic noses. I would like this to have been in the introduction.

We updated to have this line in the introduction (line 137). We did highlight the importance of sample instability and the need for the GC-MS method to determine the cut-off for measurement time in the introduction.

Line 445: Would this not be dependent on the level of infection in real life?

Indeed, the level of infection could be proportional to the bacterial concentration and directly affect VOC concentration. In this study, we used different strains from the same species and still achieved good classification.

Reviewer #3:

1. The capabilities of quantifying individual VOCs is not normally an important requirement for identification of diseases or pathogens responsible for causing diseases when using electronic-nose devices. The most important information to validate is the identities of VOCs making up the E-nose smellprint signatures and thus VOC profiles, not concentration of VOCs which is normally only needed in metabolomic studies to determine effects of pathogens on metabolic pathways of the host. Thus, quantification does not add significantly to the capabilities for testing the efficacy of a new experimental e-nose device. The ultimate objective was to develop the capabilities of the e-nose for detecting UTI caused by different microbes. The methods developed here do not contribute to that objective and the data obtained is normally part of a pilot study for methods development and not published as a stand alone unit without e-nose data on UTI samples from different types of microbial causes with adequate controls. Quantification of a single possible VOC in a UTI sample headspace provides very little information towards development of an e-nose library database containing specific complex mixtures of VOCs that affect the output signatures of the e-nose sensor array. The sensor array responds to all of the VOCs present in the headspace, not just a single VOC such as ethanol.

We appreciate the critique raised and fully understand that e-noses rely on the fingerprint of VOC profile and not necessarily concentration. However, we quantified ethanol in order to understand how sample quality is affected by measurement events and how we could maximize data collection while maintaining their quality in order to obtain a robust model as efficiently as possible. The key takeaway should be that by simply determining the proper cut-off time for data collection, we can avoid taking in unqualified data that potentially poison the model. We devised a working method for making that determination by using headspace GC-MS and disseminated a direction for future data control that helps enhance the classification accuracy.

2. All of the figures (Fig. 1-7) are of very low quality resolution and do not provide the data necessary to support the efficacy of a new experimental e-nose device for UTI diagnostics (based on quantitation of EtOH alone). Development of methods for quantitation of EtOH along with a standard curve alone do not contribute significantly towards development of e-nose methods useful for UTI diagnostics and thus the Conclusions are not supported by the objectives of the study or the data obtained towards this purpose.

We uploaded new figures of higher resolution. We agree that the GS-MS method does not directly contribute to the development of a better e-nose. However, improvement of the electronic nose involves many aspects, including choice of materials, of algorithm, and of data quality.

We explained in the introduction our rationale for choosing MOX sensors and an artificial neural network to create a potentially more powerful e-nose. We clarify in the introduction that the objectives of the study is to demonstrate the discrimination between K12 and UPEC E. coli using the EVA, which we achieved. Furthermore, ethanol concentration can be indicative of sample stability. After determining the timepoint at which the concentration deviates beyond the acceptable range, we could limit our data collection within this time frame and improve the classification model independently of either hardware (sensor choice) or software (classifier choice).

Reviewer #4:

1. The paper describes the use of an electronic nose to distinguish the pathogens causing urinary tract infection, and proposes an new experimental method to improve sample stability during e-nose measurements. The problem of sample stability during e-nose measurment should be better claimed in the introduction, because it is the focus of the described sperimentation.

We appreciate Reviewer 4 suggestion. We would like to clarify that sample stability was not improved in this study, but rather data quality was improved by terminating data collection before the samples became unstable. As suggested, we revised the introduction to better highlight the problem of sample degradation over time.

2. The methods involved to prepare EVA samples and GC-MS calibrants are described properly. However, the authors should clearly state the reasons leading to the use of normal urine samples to be inoculated rather than real infected samples.

We agree with this suggestion and revise line 379 to describe the rationale for using our method. In this paper, we devised a benchtop model of UTI by infecting the normal urine samples because it is easy to access, establish, and control in terms of bacterial culture (K12 vs UPEC) and ethanol concentration. Using real infected samples would be valuable for validating the e-nose. However, patient-derived samples could have extremely varied bacterial profiles as well as VOC profiles. Since the goal is not to simply detect between infected and non-infected, we have to carefully control the sample composition to claim the discriminating power of the e-nose. By using a benchtop model, we have decoupled confounding factors to gain a better understanding of how data quality can contribute to the robustness of our e-nose approach.

3. I think that authors should also better describe the approach involved for training the e-nose. Specifically, they should describe the scheme involved for presenting samples to the e-nose and validating the classification performance. The paragraph (lines 223-231) is not clear. The value of the sensor resistance at different times was used as features?

We included a new SI figure (S3) to show the setup of samples with the e-nose. Our feature extraction was achieved through a amplitude-driven approach of extracting the mean area under the curve for the given response duration. To focus the scope of the paper on the GC-MS validation, we did not go into much detail on this approach, however more details on this approach and the reasoning for its application can be found in our patent-filing, “Feature tuning – application dependent feature type selection for improved classification accuracy.” https://patents.google.com/patent/US20210172919A1/en

Following the preprocessing stage to extract salient features, we used a simple feed-forward multilayer-perceptron ANN consisting of three layers for classification, using the opensource GNU Octave -v5.1.0.0 (GUI).

4. All figures of the paper have a very poor quality. They should be revised. In some cases, the text is not readable. Moreover, the authors should add reference supporting their statements throughout the paper.

We revised with figures of higher resolution and more references. Please view figures individually to maintain integrity of resolution.

Submitted filename: Response to reviewers_data.docx

Click here for additional data file.

18 Jun 2022

A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.

A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.

An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols.

We look forward to receiving your revised manuscript.

Kind regards,

A. M. Abd El-Aty

Academic Editor

PLOS ONE

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.

Reviewer #1: All comments have been addressed

Reviewer #2: (No Response)

Reviewer #5: All comments have been addressed

**********

2. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: Partly

Reviewer #5: Yes

**********

3. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #5: (No Response)

**********

4. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #5: Yes

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #5: No

**********

6. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: I was excited to see your new version.

All my previous comments have been addressed.

The changes to the introduction made it more relevant and interesting to read.

On the current version of the manuscript I have noticed some grammar/editing points:

- Line 143 "in" should be added: "results (in) under two minutes".

- Line 303 there are two full stops after mL/min.

- Line 370 would be better if "there are" was used : "[...] (there are) numerous studies on the e-nose application [...]".

I hope your future research projects will go well. I am looking forward to seeing more IBM EVA use.

Reviewer #2: The paper is much improved, but there are a number of points outstanding and a number of answers were not added to the manuscript. Below are some more specific comments.

Line 97: I would add a reference to the original paper on this by Dodd & Persaud in Nature from the early 1980s.

Line 106: I am surprised there haven’t been more papers and more recent ones. These are very old. Can this table be updated?

Line 113: PCA is not a classifier, so this will need to be altered to make sense.

Line 133: I would add the paper by Estfani or urine storage by FAIMS (I might have the name wrong). It showed samples were good for 9months+ at -80C. I am also sure that some of Dutch group have done work on storage at room temperature. Might be worth adding an extra reference or two here.

Line 153: Usually ethanol is undertaken using a single optical gas sensor – for example they are used in breath ethanol testing. This should be included in the description as they are pretty common.

Line 162: I am pretty sure the EVA has been reported before. I would rephase this sentence.

Line 181: So, why not just use the ethanol sensor are be done?

Line 461: I am more interested in the stability of your instrument and the sensors over the sample. I may not have got to it yet, but how long did the calibration last?

Line 512: I would like to see a section on limitations of the study and of the use of eNoses for this purpose. For example, drift (of the various forms) is a good example. Also, how well the system will be able to cope with variations in urine. With UTIs urine can be almost clear and others the consistency of porridge.

I would still like to see a PCA from this dataset in the paper and a loading plot for the PCA. Unless the authors are saying this is the same dataset, then adding a new PCA is appropriate.

The length per test is still significant. It just needs to be added as a limitation of the study as there is no way to re-do the experiments. You can get electronic noses to respond in seconds, just not how you tested it.

From the previous points (and I state above) I can just go purchase a cheap ethanol sensor, so the reason for needing to use an electronic nose for this purpose needs to be explained in more detail. I could even get a cheap MOX sensor and make it a lot more specific to ethanol…you just need to justify better why you would use an expensive eNose over a cheap single sensor.

Details of the EVA need to make it to the paper (I didn’t see the additions). Also add details of the optimisation and not just the patent – to the paper…

Please add all feature extraction methods to the paper. Also software used etc.

Reviewer #5: Though the manuscript is well designed and presented, I am doubtful regarding the applicability of the developed method. The manuscript can be considered for publication provided that comments and suggestions given by the other reviewers are fully addressed.

**********

7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

Reviewer #5: No

**********

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step.

30 Jul 2022

RESPONSE TO REVIEWERS (PONE-D-21-22287)

We are delighted to receive positive comments from reviewers # 1 and 5 and acknowledge concerns raised by reviewer #2. We have addressed all the comments in our revised manuscript, and the response details are as follows. We are happy to answer further concerns and questions from the reviewers.

Reviewer 1:

"I was excited to see your new version. All my previous comments have been addressed. The changes to the introduction made it more relevant and interesting to read. I hope your future research projects will go well. I am looking forward to seeing more IBM EVA use."

We appreciate all the positive comments from Reviewer 1 and grateful for your careful evaluation of our manuscript.

Line 143 "in" should be added: "results (in) under two minutes"

Please note that the line numbers hereafter refer to the no-mark-up version of our new revision.

Line 148: We corrected it.

Line 303 there are two full stops after mL/min.

I could not identify the mistake. The manuscript had only one full stop after mL/min at line 301.

Line 370 would be better if "there are" was used : "[...] (there are) numerous studies on the e-nose application [...]".

Thank you for your suggestion. We made the change at line 379.

Reviewer 2

"The paper is much improved, but there are a number of points outstanding and a number of answers were not added to the manuscript. Below are some more specific comments."

Line 97: I would add a reference to the original paper on this by Dodd & Persaud in Nature from the early 1980s.

Please note that the line numbers hereafter refer to the no-mark-up version of our new revision.

Line 87: We added the paper, which was an excellent reference to the origin of e-nose technology.

Line 106: I am surprised there haven't been more papers and more recent ones. These are very old. Can this table be updated?

Line 95: We updated the table with a few more recent references. However, to our surprise, not many new papers specifically looked at the urine-based classification of pathogens in the context of urinary disease. We reflected this in the paragraphs following table 1.

Line 113: PCA is not a classifier, so this will need to be altered to make sense.

We removed that line and modified the paragraph (not lines 97-108).

Line 133: I would add the paper by Estfani or urine storage by FAIMS (I might have the name wrong). It showed samples were good for 9months+ at -80C. I am also sure that some of Dutch group have done work on storage at room temperature. Might be worth adding an extra reference or two here.

Line 114: We appreciate your helpful comment. Esfahani et al. (2016 & 2018) are good papers to be included as it shows that despite storing at -80oC, storage time still significantly affects classification accuracy. They suggested samples were good within 9 months (in the 2016 paper) and that performing urine analysis within a year of storage resulted in better accuracy than including all samples within four years. Their 2018 study only considered two labels (healthy vs. diabetes). An even shorter storage time could be recommended for a more complex analysis involving more labels.

Line 153: Usually ethanol is undertaken using a single optical gas sensor – for example they are used in breath ethanol testing. This should be included in the description as they are pretty common.

Line 138: We updated this line to add the breath analyzer as an example of an ethanol quantification system.

Line 162: I am pretty sure the EVA has been reported before. I would rephase this sentence.

Line 146 We have introduced the EVA before in a conference proceeding (Adebiyi et al. 2019). Therefore, we rewrote the sentence as "In this paper, we used the Electronic Volatile Analyzer (EVA)…" and cited this reference.

Line 181: So, why not just use the ethanol sensor are be done?

Line165: Ethanol sensors alone will not create enough differential response to discriminate many different VOC profiles. We mainly measured ethanol as an indication of sample stability because we inoculated urine with ethanol-producing bacteria (E. coli). However, a robust e-nose requires a combination of sensors responsive to various compounds. We revised this line to reflect these points.

Line 461: I am more interested in the stability of your instrument and the sensors over the sample. I may not have got to it yet, but how long did the calibration last?

Line 434: We believe we shared this in our previous review, but the stability of our sensors over the sample period was reproducible from measurement to measurement. From a cold-start of the instrument, we ran air (our carrier gas) through the chamber for sixty minutes, before we moved on to the measurement samples. Between each 15 minute measurement, we ran air for five minutes to purge any residual VOCs.

Line 512: I would like to see a section on limitations of the study and of the use of eNoses for this purpose. For example, drift (of the various forms) is a good example. Also, how well the system will be able to cope with variations in urine. With UTIs urine can be almost clear and others the consistency of porridge.

Line 481: We added a paragraph on the limitations of the study. We think the consistency of urine is not a problem per se but what VOC profile they will produce is the determining factor. Obviously, cloudy vs clear urine may have different VOCs in them, some may not be even related to UTI. Therefore, screening for prevalent compounds in UTI vs healthy controls is important and optimizing sensor choice that targets these compounds. Sensor drift is a notable limitation of electronic noses given the variability of the MoX sensors in undergoing these oxidation reactions over time. Sensor drifting can be compensated by a multitude of mathematical techniques (newly included in the paper) such as baseline correction and the like, but our instrument is designed for real-time use”in the wild”, so this choice of calibration was a design choice meant to be as robust as possible to accommodate mostly operator-free scenarios.

I would still like to see a PCA from this dataset in the paper and a loading plot for the PCA. Unless the authors are saying this is the same dataset, then adding a new PCA is appropriate.

Our dataset is the same as the last review. We have tried PCA but found ANNs more suitable to capture the drift and variability in the model for our use-case and setup (which we also described previously). We are happy to provide the reviewer with a PCA for their reference, but a study by itself can be conducted on the choice of mathematical approaches suitable. Given the scope of this paper was mainly to quantify the levels of ethanol in our samples with the validated GC-MS method, we did not go into all those details. An example of a reference related to these mathematical methods with the PCA loadings can be found in Figure 3 of this reference [30] in our paper, which provides these details.

Adebiyi, A., Than, N., Swaminathan, S., Abdi, M., Bowers, A. N., Fasoli, A., ... & Bozano, L. (2020, January). Rapid Strain Differentiation of E. coli-inoculated Urine Using Olfactory-based Smart Sensors. In SEIA'2019 Conference Proceedings (p. 307).

The length per test is still significant. It just needs to be added as a limitation of the study as there is no way to re-do the experiments. You can get electronic noses to respond in seconds, just not how you tested it.

We have addressed it in the limitation paragraph. However, we would like to add that this length (15 minutes) signifies the measurement stage to train the model. At inference, classification takes 80 seconds.

From the previous points (and I state above) I can just go purchase a cheap ethanol sensor, so the reason for needing to use an electronic nose for this purpose needs to be explained in more detail. I could even get a cheap MOX sensor and make it a lot more specific to ethanol…you just need to justify better why you would use an expensive eNose over a cheap single sensor.

Line 475: We appreciate that cheap ethanol sensors are ubiquitous and finding a correlation between the quantity of ethanol and E. coli strain type could potentially aid in strain differentiation. In some ways, this is what we sought to explore by using the validated GC-MS method to determine if the change in ethanol is mainly due to the degradation of the sample vs the ethanol secreted by the given E. coli strain at incubation. It could be more difficult for a cheap ethanol sensor to tell if it is UPEC E. coli after the sample had been sitting for three days than K12 after two. If we did seek to use the cheap ethanol sensors, we would stil have to validate it, but we could add that as a point. Speaking to the justification of the EVA as a device, we are seeking to build a library of different types of bacteria, not just speficially those that are related to ethanol, it just happens that ethanol is the main VoC in this use-case.

Details of the EVA need to make it to the paper (I didn't see the additions). Also add details of the optimisation and not just the patent – to the paper…

As we mentioned previously, the scope of this paper is intended to look at the quantification of the ethanol method, and not to provide a deep dive into the hardware optimization of the EVA, which could be its own study. We are happy to reference this patent in paper, but believe that adding all the details will take away from the focus of this study. However, we briefly summarize the operational basis of EVA on page 10, which we believe is appropriate to reflect the focus of this paper.

Please add all feature extraction methods to the paper. Also software used etc.

Line 247: We have added this to the paper.

Reviewer #5:

Though the manuscript is well designed and presented, I am doubtful regarding the applicability of the developed method. The manuscript can be considered for publication provided that comments and suggestions given by the other reviewers are fully addressed.

We are thankful for the positive comment. We believe other researchers in the field can adapt the developed method to characterize the content in their e-nose samples. Not only for ethanol but other compounds can also be quantified by modifying the validated method.

Submitted filename: Response to Reviewers_July2022review.docx

Click here for additional data file.

20 Sep 2022

Quantitation of ethanol in UTI assay for volatile organic compound detection by electronic nose using the validated headspace GC-MS method

PONE-D-21-22287R2

Dear Dr. Adebiyi,

We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements.

Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication.

An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at http://www.editorialmanager.com/pone/, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to-date. If you have any billing related questions, please contact our Author Billing department directly at authorbilling@plos.org.

If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

Kind regards,

A. M. Abd El-Aty

Academic Editor

PLOS ONE

Additional Editor Comments (optional):

The authors respond satisfactorily to the comments raised by the reviewer.

Reviewers' comments:

27 Sep 2022

PONE-D-21-22287R2

Quantitation of ethanol in UTI assay for volatile organic compound detection by electronic nose using the validated headspace GC-MS method

Dear Dr. Adebiyi:

I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department.

If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact onepress@plos.org.

If we can help with anything else, please email us at plosone@plos.org.

Thank you for submitting your work to PLOS ONE and supporting open access.

Kind regards,

PLOS ONE Editorial Office Staff

on behalf of

Prof. A. M. Abd El-Aty

Academic Editor

PLOS ONE