Rapid detection of SARS-CoV-2 infection by multicapillary column coupled ion mobility spectrometry (MCC-IMS) of breath. A proof of concept study

There is an urgent need for screening of patients with a communicable viral disease to cut infection chains. Recently, we demonstrated that ion mobility spectrometry coupled with a multicapillary column (MCC-IMS) is able to identify influenza-A infections in patients’ breath. With a decreasing influenza epidemic and upcoming SARS-CoV-2 infections we proceeded further and analyzed patients with suspected SARS-CoV-2 infections. In this study, the nasal breath of 75 patients (34 male, 41 female, aged 64.4 ± 15.4 years) was investigated by MCC-IMS for viral infections. Fourteen were positively diagnosed with influenza-A infection and sixteen with SARS-CoV-2 by reverse transcription polymerase chain reaction (RT-PCR) of nasopharyngeal swabs. In one patient RT-PCR was highly suspicious of SARS-CoV-2 but initially inconclusive. The remaining 44 patients served as controls. Breath fingerprints for specific infections were assessed by a combination of cluster analysis and multivariate statistics. There were no significant differences in gender or age according to the groups. In the cross validation of the discriminant analysis 72 of the 74 clearly defined patients could be correctly classified to the respective group. Even the inconclusive patient could be mapped to the SARS-CoV-2 group by applying the discrimination functions. Conclusion: SARS-CoV-2 infection and influenza-A infection can be detected with the help of MCC-IMS in breath in this pilot study. As this method provides a fast non-invasive diagnosis it should be further developed in a larger cohort for screening of communicable viral diseases. A validation study is ongoing during the second wave of COVID-19. Trial registration: ClinicalTrial.gov, NCT04282135 Registered 20 February 2020—Retrospectively registered, https://clinicaltrials.gov/ct2/show/NCT04282135?term=IMS&draw=2&rank=1


Introduction
To interrupt infection chains in SARS-CoV-2 disease, screening methods are urgently needed. The diagnostic standard is reverse transcription polymerase chain reaction (RT-PCR), but deep nasopharyngeal swabs taken by trained personnel are required and rapid PCR techniques still take more than 30 min [1]. Particularly for screening at airports or other sites where rapid screening of asymptomatic patients is demanded, this method is logistically challenging and usually takes far longer than 30 min. Additionally, mainly within the scope of developing countries, RT-PCR is far too expensive. So, there is still a need for a faster, really non-invasive screening tool [2]. A screening tool that can be easily used and prevents false negative results would fulfill this demand. Even with a lower specificity than PCR, such a method could be used as a pre-screening tool with application of PCR in the positive patients, decreasing the number and financial impact of PCRs necessary.
In classic antiquity, without other diagnostic tools physicians had to rely on their basic senses: seeing, touching, hearing and smelling [3] With improved technical possibilities, these skills have been moved into the background. Physicians are only taught a few scents in medical school, like acetone for ketoacidosis and ammonia for liver disease. It is well known that different bacteria smell differently. Pseudomonas aeruginosa has a linden blossom scent, while Escherichia coli smells fecal. Based on their better olfactory senses animals have been trained to smell infectious diseases but lack of reproducibility precludes wider application [4].
The scents of infectious diseases are volatile organic compounds (VOCs) that are released by the metabolism of the germ or the host. There are different technical approaches to discriminate pathogens or diseases based on VOCs but none have been used regularly in clinical practice [5].
Compared to bacteria, viruses have no own metabolism. So, the scents can only relate to the host response of the viral infection [10]. GC coupled mass spectrometry (GC-MS) was able to detect influenza infection of cell cultures in vitro [11].
As GC-MS is not feasible for point-of-care diagnostics there are attempts to train dogs to sniff for viral diseases, currently for SARS-CoV-2 associated disease [12,13]. In a recently published study dogs were trained to smell respiratory samples from COVID-19 patients with an average detection rate of 94% [14].
In a recent study we demonstrated that influenza-A infection can also be detected in the breath of influenza-A infected patients by multicapillary column coupled ion mobility spectrometry (MCC-IMS) [15]. Therefore, we extended this study to analyze whether MCC-IMS is also able to detect SARS-CoV-2 infection in breath.

Patients
During the influenza-A epidemic 2020 [15] 11 males, 13 females, aged 66 ± 14.2 years and, during the SARS-CoV-2-pandemic, 23 males, 28 females, aged 63 ± 16 years with suspected infections, including four healthy staff members, were asked to participate in the study. The study was approved by the Ethics Committee of Erlangen University #426_18_B and registered at ClinicalTrials.gov (NCT04282135).
After routine RT-PCR analysis of nasopharyngeal swabs 14 of the patients were positive for influenza-A, 16 were positive for SARS-CoV-2, and the remaining 40 patients and four healthy staff members were used as controls. In one patient SARS-CoV-2 PCR was inconclusive. This patient is omitted in the statistical analysis leading to three groups: influenza-A infected, SARS-CoV-2 infected and not-infected controls.
Breath samples were taken and analyzed by MCC-IMS. Patients were recruited between 8 April 2020 and 7 May 2020. As time went on, measures of social distancing and segregation were successful leading to decreasing numbers of SARS-CoV-2 patients.

PCR
SARS-CoV-2 was tested by taking a deep nasopharyngeal swab by applying the 'Xpert® Nasopharyngeal Sample Collection Kit for Viruses' (Cepheid, Maurens-Scopont, France) and performing real time PCR by applying the Allplex 2019-nCoV Assay (Seegene, Seoul, South Korea) on the CFX 96 Real-Time System (BioRad, Feldkirchen, Germany) after extracting RNA using the StarMag 96 UniTube Kit (Seegene) on the SGPrep32 extraction system (Seegene). Due to shortage of supply, RNA was in part alternatively extracted by applying the QiaAmp DSP Virus Spin Kit (Qiagen, Hilden, Germany) using the Qiacube automated extraction system (Qiagen). Influenza PCR was performed with the 'Xpert® Xpress Flu/RSV' (Cepheid) on the 'Infinity' (Cepheid).

Breath sampling and MCC-IMS
For the IMS we used the MCC-IMS-device from STEP Sensortechnik und Elektronik, Pockau, Germany (STEP IMS NOO). The device is distributed as a medical device (In-vitro-diagnostic) in combination with evaluation software as the 'MultiMarkerMon-itor®' by Graupner medical solutions GmbH, Geyer, Germany.
All patients were connected by a foam-cuffed oxygen catheter (#01442958, Asid-Bonz, Herrenberg, Germany) via a 0.22 µm filter (Navigator Lab Instruments, Tinajin, China) and a Perfusor Line (B Braun, Melsungen, Germany) directly to the MCC-IMS. Patients were instructed to take a deep breath and to exhale slowly through the nose. During the exhalation breath was sampled for 10 s.
The STEP device directly draws the sample by an internal pump (200 ml min −1 ) into the analyzing circuit without any pre-analytical procedures. A pump fills the loop made of perfluoroalkoxy polymer with a flow of 200 ml min −1 . The time of the filling process is adjustable in increments of seconds. It must be ensured that the loop is filled with the sample. During the filling time the sample is drawn through the loop and leaves the device by the waste gate. It is ensured that the loop is filled with 2 ml of the sample after stopping the pump. These 2 ml are let into the MCC column by a valve. Then, the sample is pre-separated by an isothermally heated multicapillary GC column (60 • C) into single analytes, which enter the IMS unit based on their retention times. In the IMS unit the analytes are ionized by beta radiation of a tritium source below the free limit for radiation (99 MBq). Afterward, the generated ions are accelerated in a 50 mm long drift-tube under the influence of an electric field (400 V cm −1 ) toward the detector, which is also tempered to 60 • C. On their way the positive ions collide with air molecules from the drift gas flowing in the opposite direction, and are separated depending on their ion mobilities and detected by the collector electrode sampled every 10 µs. The ion gate has an adjustable opening time (currently 100 µs) and is pulsed every 30 ms. There are 16 single spectra per second that are averaged and undergo a wavelet transformation. The denoised spectrum is further used for analysis. The received IMS spectra are stored internally in the device and later analyzed offline.
The used IMS device is equipped with a circulation filter and internal gas circulation. Using a circulation pump, ambient air filtered by an activated carbon filter was provided as drift gas (400 ml min −1 ) and carrier gas (20 ml min −1 ) to the device. There are different IMS and field-asymmetric-IMS devices available. Most of these devices need an external gas cylinder to support the device with synthetic air or nitrogen as drift and carrier gases. The STEP device uses filtered air and does not need an external gas supply.
Quality control measures have been previously described [15].

Data analysis
The VOCs are characterized by their retention time in the MCC and the drift time in the IMS.
One spectrum over 2048 measurement points every 10 µs (in total 20.48 ms) is obtained every second for a total time of 240 s.
These spectra can be visualized on a heatmap with retention time on the y-axis and the drift time on the x-axis. For an example of a heatmap please refer to the online supplement.
To decrease the complexity of the data we used a proprietary cluster analysis software using a support vector machine (European Patent EP 2 729 801 B1) [16]. The principle of this cluster analysis software has also been described before [8]. In brief, after baseline correction for noise the software determines the peaks of each measurement based on the intensity signal threshold and categorizes them by retention time and drift time. Depending on these parameters the clusters are numbered assuming that every cluster represents a distinct VOC. Peaks of different measurements with similar drift and retention times according to a defined threshold are mapped to the same clusters. For the parameters of the cluster analysis please refer to table S2 in the supplement (available online at stacks.iop.org/JBR/15/027105/mmedia).

Statistical analysis
Due to the small sample size and the lack of normal distribution the Mann-Whitney U test and Kruskal-Wallis test were applied for differences in patient characteristics. Patients from both sub-studies, where neither influenza-A nor SARS-CoV-2 were found in the PCR, were combined as controls for the combined dataset.
To exclude cross-correlated clusters we performed a stepwise canonical discriminant analysis for optimal minimization of Wilks lambda. To enter or remove variables from the model F significances of 0.05 and 0.1 were used.
For the statistical analysis we used IBM SPSS 22.0 (IBM, Armonk, NY). The details of the statistics are listed in the supplement.

Results
Age and infection markers as well as biometrics were not significantly different between the groups. The only significant difference was the reduced leukocyte count in the SARS-CoV-2 positive population (table 1). The influenza-A positive patients were included to verify that the identified specific clusters of SARS-CoV-2 did not simply reflect metabolic changes of viral infection.
Although the patients were triaged as suspected by the symptoms at presentation there was significantly more fever, coughs and dyspnea in the SARS-CoV-2 positive group as well as in the influenza-A group. Only gastrointestinal symptoms did not differ between the groups. Compared to influenza, dyspnea was significantly more common in the SARS-CoV-2 group (table 2 and figure 1).
In one patient SARS-CoV-2-PCR was initially inconclusive. But as the manufacturer (Seegene) meanwhile redefined the definition for positive test results this patient would now be categorized as positive. This result was reproducible in a second test.
One hundred and fifty-five clusters were found that were further used for the multivariate analysis  (table S5 in the supplement). Thirty-two clusters were needed in the stepwise approach to minimize Wilks lambda (table S6 in the supplement). As no validation cohort was available, cross validation using the leave-one-out method was employed. By applying cross validation 97.3% of the cases were correctly classified by the two derived canonical discrimination functions, even between influenza-A and SARS-CoV-2. There was only one control misclassified as influenza-A and one as SARS-CoV-2, respectively. There were no false negatives in this analysis (table 3).
Discriminant analysis is able to explain 100% of the variance. In the scatter plot of the two discriminant functions all three groups are separated nicely (figure 2). The patient with the initial inconclusive PCR result had low viral load and is nevertheless mapped to the SARS-CoV-2 group. For this patient one RT-PCR showed only a signal of the E-as well as N-gene with a PCR crossing point of 35.44 for both genes. A second PCR performed two days later gave a positive signal for the RdRP-gene with a crossing point of 37.65, while the E-and N-gene were then negative.
Another validation was performed by dividing the entire dataset into a training subset (35 controls, 11 influenza-A and 13 SARS-CoV-2 patients) and a validation subset (nine controls, three influenza-A and three SARS-CoV-2 patients) using the Kennard Stone algorithm.
The model was calculated based on the training set. Then, the scoring wizard was used to predict the validation subset. Accuracy was 12/15 (80%, 95% confidence interval 52%-96%). Again, no influenza-A or SARS-CoV-2 patients were wrongly predicted as controls (table 4).

Discussion
Today RT-PCR is considered the gold standard in the diagnosis of SARS-CoV-2 infection, but even in trained hands a false negative rate of about 25% and   a false positive rate of 2.3%-6.9% has to be expected [17]. However, false negative results are often caused by low swab and sampling technique quality, while in this study only well-trained staff performed the sampling by applying brushed swabs from Copan, which can be considered as high quality. Hence, we only noticed false negative nasopharyngeal swabs in patients who were either in the pneumonia phase of COVID-19 or were at the detection limit.
Our clinical experience was that patients at the detection limit are usually at the very end of a COVID-19 infection and are most likely not infective anymore [18]. As our patients were examined at admission this effect can be neglected in our study.
Currently SARS-CoV-2 infections are rising in several countries causing tremendous cost. In underdeveloped countries neither trained staff nor financial power to afford RT-PCR for mass screening is available [2]. Hence, the MCC-IMS could provide an alternative testing perspective for these countries.
To the best of our knowledge our study is the first showing that breath-analysis is able to discriminate SARS-CoV-2 infected patients from controls with respiratory infections and influenza-A infection as well. Even one suspected but not clearly PCR-confirmed patient could be assigned to the SARS-CoV-2 group and turned out to be positive later.
The study by Ruszkiewicz et al showed that patients with COVID-19 could be distinguished from non-infected patients by GC-IMS analysis of breath. There, specific markers were detected in elevated concentrations in the breath of SARS-CoV-2 infected patients compared to patients without this infection [19].
We used MCC-IMS because of the ease of application. The STEP-IMS device does not need any preanalytic procedures or external gases. So, no shortage of swabs, tubes or reagents has to be faced in the scope of a pandemic.
The device draws the breath into the system by an internal pump. This simplifies the sampling compared to other IMS devices, where absorption/desorption tubes are needed [9].
It may be argued that due to the automatic suction a dilution by room air may be possible. The only dilution taking place is from the air in the dead space of the tubing, which is about 6 ml. In our setting the foam cuff seals the nostril from the room air. Besides the air from the pulmonary compartment, the only air that is drawn into the system is the air in the pharyngeal space that is assumed to be also infected in SARS-CoV-2 infection. The volume of this space is exceeding the drawn 33 ml. We chose this volume to reduce the amount of dilution by the dead space.
A possible optimization could be the use of tubing with less dead space and reduction of the volume of air drawn.
The only task to be fulfilled by less-trained staff is to introduce the foam-cuffed catheter into a nostril of a spontaneously breathing individual and to hit a key to start the measurement. In the study by Ruszkiewicz et al the patients had to exhale through a conicalshaped tube, where the breath had to be sampled by hand with a 5 ml syringe to be injected into the IMS device [19]. Compared to the direct aspiration of the sample by the device used in our study, such a procedure needs compliant patients as well as trained staff and may not be feasible for large scale screening. Furthermore, a dilution of the exhaled breath with ambient air might lower the sensitivity of this preanalytical procedure.
Compared to the above-mentioned study our approach is also applicable to uncooperative, e.g. delirious or demented, patients; albeit such patients were excluded in our study as they would not be able to actively consent.
It could be argued that IMS only provides peaks according to retention time and drift time, while MS is able to chemically describe the VOCs. But also in MS not all peaks are clearly assigned to a chemical substance, and some are therefore also only numbered or characterized by the time of flight [6]. Furthermore, even though it is interesting to chemically identify the relevant peaks in the breath of COVID-19 patients it seems to be dispensable for SARS-CoV-2 screening.
In the Ruszkiewicz study the different substances have been identified as ethanal, acetone, acetone/2butanone cluster, octanal, Feature 144, isoprene, heptanal and propanal, but the metabolic origin of these substances cannot be described either. So, compared to the unnamed clusters in our study this is a pattern of named VOCs, which of course adds to the understanding of the method. However, medication, food and room air might differ between patients and, under certain circumstances, interfere with peaks used for the identification of SARS-CoV-2 positive persons. This interference is present no matter whether the VOCs are identified or not, and it is extremely difficult if not impossible to predict which combination of food, medication, metabolic state and room air might interfere with the relevant peaks taken for the diagnosis 'SARS-CoV-2 positive' , and this holds true for the study of Ruszkiewicz et al as it does for ours.
It has been shown that dogs can discriminate sputum from SARS-CoV-2 infected individuals [13,14]. In some countries like Canada scent dogs are trained for SARS-CoV-2 screening. The scents of sputum or breath are a composition of different VOCs. So, a discrimination by VOC patterns in a noise of medication, food, etc seems to be possible. Hence, we think that not knowing the exact chemical structure of the VOC is no detriment. Besides statistical methods, we are starting to investigate if pattern recognition by artificial intelligence can improve the discrimination.
As the scent of the breath does not rely on the virus itself but on the host response to the infection, cross-reactivity of breath analysis with other viral infections has to be expected [10]. As MCC-IMS could differentiate between SARS-CoV-2 and influenza-A infection we assume that at least to a certain extent different viruses may cause different host responses and, therefore, produce different fingerprints of IMS spectra. However, this needs to be addressed in future studies.
Similar to antibody tests an overlap of the VOCs with other corona virus infections has to be anticipated. But this is a constraint every analysis of metabolomics has to face. However, within the current pandemic we detected almost no endemic corona viruses in our adult patients by multiplex PCR.
Compared to other breath analysis studies we did not require fasting before the sampling. Although fasting state may reduce interferences with other metabolism, it will not be feasible for large scale screening.
Another necessity for further progress of breath analysis in screening for infections is the extension of the study to other ethnicities and civilizations to investigate whether ethnos and lifestyle need to be considered for the analysis.
One drawback of our study is the limited number of patients. Many patients with SARS-CoV-2 were not able to give informed consent. Another point is the weakening of the SARS-CoV-2 wave in Germany at the end of April as this led to a slowing of accrual.
We therefore assess this study as a proof of concept and encourage other researchers to further investigate breath analysis by MCC-IMS for the detection of SARS-CoV-2 infections. We are currently working on a point-of-care prototype with an instant analysis of the data as this will be the relevant step for large scale screening.
As MCC-IMS is fast, non-invasive and does not need any reagents or pre-analytical procedures it seems promising for a screening or at least prescreening device, even in underdeveloped countries or air travel. In summary, we identified a quick and cheap means of large-scale SARS-CoV-2 testing.

Conclusion
In our small cohort analysis the use of MCC-IMS is able to discriminate between influenza-A, SARS-CoV-2 infections and controls in a few minutes.
As this method is completely non-invasive and does not need any reagents or pre-analytic procedures it seems promising as a mass-screening device, even in underdeveloped countries.
We encourage further trials to use this technique in different patient settings.

Data availability statement
The data that support the findings of this study are available upon reasonable request from the authors.