Breath analysis by ultra-sensitive broadband laser spectroscopy detects SARS-CoV-2 infection

This study was approved by the Institutional Review Board (protocol no. 21-0088) of the University of Colorado Boulder. From May 2021 to January 2022, breath samples from a total of 170 research subjects were collected with a class distribution for SARS-CoV-2 infection of 83 positives (48.8%) and 87 negatives (51.2%). Research subjects were all University of Colorado Boulder affiliates, at least 18 years old, and recruited after taking a university-provided saliva-based or nasal swab COVID-19 RT-PCR test. The general campus population was $\gt$90% vaccinated. No participants were severely ill or requiring hospitalization at the time of their sample collection. After receiving their COVID-19 test results, potential subjects received a study recruitment email and were asked to contact the research team within 24 h if interested in participation. They then reviewed and signed an informed consent form, completed a questionnaire, and scheduled an appointment for the collection of their breath samples. The questionnaire collected self-reported information on sex, age, and race as well as other factors that could impact breath analysis including smoking, alcohol use, and underlying gastrointestinal symptoms. Additional information was collected on acute symptoms experienced by the positive participants. No viral genomes were sequenced, but the Colorado statewide data [22] over our subject recruitment period indicates infection with several viral variants associated with several infection waves (namely, alpha, delta, and omicron) in the community. All data (i.e. informed consent form, questionnaire, and Tedlar bag ID) were collected and managed using the REDCap electronic data capture tool [23, 24] hosted by the University of Colorado Denver.

Standard Tedlar bags (1 l, part no. 249-01-PP, SKC Inc.) were used to collect exhaled breath. During the sample collection appointment, research subjects were asked to hold their nose and breathe through their mouth. They were instructed to inhale to full lung capacity for 1–3 s, followed by exhaling the first half of their breath to the surroundings and the second half into the bag until the latter was above ∼80% full. The sample collection location was an outdoor university parking lot. The participants were not instructed to limit or control their smoking, food or alcohol intake prior to sample collection. Right after collection of one breath sample, the Tedlar bag was stored inside an air-tight container at ambient temperature and transported to the indoor lab housing the CE-DFCS setup for immediate data collection and analysis. The breath sample was warmed to 37 $^{\circ}\mathrm{C}$ for 20 min to reduce condensation, then steadily flowed through the cleaned vacuum chamber held at room temperature (20 $^{\circ}\mathrm{C}$) at a rate of ∼1 l min⁻¹. Just before bag exhaustion, timely closure of the gas valves detained a portion of breath sample inside the chamber and a static pressure of 50 Torr (67 mbar) was reached (without re-condensation) for spectroscopic data collection. After the measurement, the breath sample was pumped out to an exhaust line leading to the building exterior. The used Tedlar bag was autoclaved and disposed of. While direct sampling at atmospheric pressure by our breathalyzer is feasible, off-line sampling and negative pressure were adopted to ensure no SARS-CoV-2 could be introduced into the laboratory air. Spectroscopy data collection for each breath sample was completed in less than 10 min. This can be further reduced to about 1 s when optimized data acquisition and readout are implemented. Overall, from sample collection and transportation to completion of data analysis, the total time was less than an hour. Air samples were collected on separate days over the subject's recruitment period at the sample collection location as control specimens.

The working principle of the CE-DFCS breathalyzer is illustrated in figure 1(a). A high-resolution broadband absorption spectrum, consisting of a total of 14 836 distinct molecular features each measured ultra-sensitively at individual optical frequencies, was recorded for each breath sample (see sample spectrum in figure 1(b)). The breath spectrum was processed by machine learning analysis for binary response classifications. For additional instrument details, see [19].

Zoom In Zoom Out Reset image size

We employed two spectral pre-processing techniques for machine learning analysis: (1) a pattern-based approach that directly used all 14 836 molecular absorption features as the predictor variables; (2) a molecule-based approach that used 16 known small molecule compounds (H₂O, HDO, ¹²CH₄, ¹³CH₄, OCS, C₂H₄, CS₂, H₂CO, NH₃, CH₃OH, O₃, N₂O, NO₂, SO₃, HCl, and C₂H₆) fitted to the spectra as predictor variables. The former approach identifies all stable patterns that can be used for diagnostics, whereas the latter identifies only the patterns that can be reduced to known molecular identities, which may result in loss of utilizable chemical information but allows better interpretability into the model details. The 16 compounds were chosen due to their availability from the high-resolution transmission molecular absorption database [25]. While more molecules can potentially be uncovered and fitted, quantitative extraction of their identities requires cross-sectional data at our experimental conditions (20 $^{\circ}\mathrm{C}$ temperature and 50 Torr pressure) to be available. Unfitted species are hence not used in the molecule-based analysis despite being potentially useful to facilitate better predictive power.

To enable binary class assignment, we used partial least squares-discriminant analysis (PLS-DA) [26]. This method allows for the reduction of high-dimensionality data into a one-dimensional scalar number to differentiate between the opposing response classes (positive vs. negative). Variable importance in the projection (VIP) scores [27] were determined for assessing the relative importance of each predictor variable. To assess predictive power, the complete dataset (N = 170) was randomly divided into a training set (n = 140) and the remaining as a testing set (n = 30). Both sets shared the same binary class distributions as the complete data set. The training set was used for model construction (a total of 15 PLS components were constructed) and the testing set was used for a blind test to obtain a receiver-operating-characteristic (ROC) curve, from which the area under the curve (AUC) value was calculated. Depending on how the complete set was divided, the AUC value obtained can vary to a certain extent. To ensure convergence, we repeated the whole process (i.e. cross-validation) for a total of 10 000 times, and each time a new training set and testing set were randomly re-selected for a new AUC value to be calculated. The ROC curves generated from the total of 10 000 cross-validation runs were averaged together to obtain an averaged ROC curve. The AUC of the averaged curve thus represents the average AUC from all cross-validation runs. To determine the AUC uncertainty, we used different training/testing partition ratios and different numbers of PLS components. All analysis code was written using MATLAB, and the PLS-DA was performed using the built-in package based on the SIMPLS algorithm [28]. The supplementary file contains additional details on PLS-DA and VIP score principles, ROC averaging, and AUC uncertainties.

One-hundred and seventy participants enrolled in this study, with characteristics summarized in table 1. These included 83 (48.9%) SARS-CoV-2 positive subjects and 87 (51.2%) SARS-CoV-2 negative subjects based on prior RT-PCR tests. The median age was 22 years in the infection-positive and 24 years in the infection-negative groups (p < 0.05). Both infection-positive and negative groups were balanced for sex (53.0% female infection-positives, 49.4% female negatives). Race and ethnicity distributions were equivalent between infection-positive and negative groups. A higher number of infection-negative subjects reported a history of rare to occasional abdominal symptoms, though there was no difference in the history of lactose intolerance or constipation between the two groups. SARS-CoV-2-positive subjects were asked additional questions regarding COVID-19-related symptoms, if any (table 2). We found most subjects reported multiple symptoms (figure 2). Of 78 who responded, 50.0% reported 5–7 of the 11 listed symptoms, 5.1% were asymptomatic, and 2.6% reported 10 symptoms.

Zoom In Zoom Out Reset image size

Table 1. Participant characteristics.

Characteristic	Total (N = 170)	SARS-CoV-2 positive (n = 83; 48.9%)	SARS-CoV-2 negative (n = 87; 51.2%)	P a
Sex
Female	87 (51.2)	44 (53.0)	43 (49.4)	0.99
Male	83 (48.8)	39 (47.0)	44 (50.6)
Age, median (IQR), years	23 (8.8)	22 (6)	24 (10)	0.01
Race
Other/mix	12 (7.0)	4 (4.8)	8 (9.2)	0.53
Asian	20 (11.8)	8 (9.6)	12 (13.8)
White	138 (81.1)	71 (85.5)	67 (77.0)
Latino
Yes	14 (8.2)	7 (8.4)	7 (8.0)	0.84
No	156 (91.8)	76 (91.6)	80 (92.0)
Alcohol frequency, days week−1
d = 0	45 (26.5)	15 (18.1)	30 (34.5)
$0 \lt d \leqslant 3$	114 (67.1)	64 (77.1)	50 (57.5)	0.09
$3 \lt d \leqslant 7$	11 (6.5)	4 (4.8)	7 (8.0)
Smoker
(Tobacco/Vape/Marijuana)
Yes	31 (18.2)	11 (13.3)	20 (23.0)	0.05
No	139 (81.8)	72 (86.7)	67 (77.0)
Abdominal pain
Never	79 (46.5)	48 (57.8)	31 (35.6)	0.01
Rarely	50 (29.4)	20 (24.1)	30 (34.5)
$\geqslant$Occasionally	41 (24.1)	15 (18.1)	26 (29.9)
Lactose intolerance
Not at all	113 (66.5)	60 (72.3)	53 (60.9)	0.15
Very mild to mild	34 (20.0)	14 (16.9)	20 (23.0)
Moderate to severe	23 (13.5)	9 (10.8)	14 (16.1)
Constipation
Not at all	140 (82.4)	69 (83.1)	71 (81.6)	0.82
Very mild	19 (11.2)	9 (10.8)	10 (11.5)
$\geqslant$Mild	11 (6.4)	5 (6.0)	6 (6.9)

Information collected for the total of N = 170 participants (n = 83 positive; n = 87 negative). Unless otherwise indicated, data are presented as n (%). IQR, interquartile range. ^a P values compare subjects positive and negative for SARS-CoV-2 infection.

Breath analysis by laser spectroscopy can differentiate between SARS-CoV-2 infection positives and negatives. Using the two spectral pre-processing techniques for machine learning analysis, we found the pattern-based approach yielded an AUC of 0.849 (standard deviation [SD], 0.004) (figure 3(b)) and the molecule-based approach yielded an AUC of 0.769 (SD, 0.007) (figure 3(e)). Both approaches confirmed that significant differences in breath contents caused by SARS-CoV-2 infection was successfully detected by CE-DFCS. The classification results on SARS-CoV-2 infection should be interpreted as the co-agreement between the CE-DFCS breath test and the RT-PCR tests employed. As control experiments to validate the analysis methodology, we checked predictions for two cases with known responses: (1) a random guess based on subjects born in even vs. odd months, for which the lowest possible AUC of 0.5 is expected; (2) a perfect discrimination comparing ambient air vs. exhaled breath samples, for which one expects an AUC of 1. Both the pattern-based and molecule-based approaches confirmed expectations for results from a random sampling by birth month (figures 3(a) and (d)), yielding an AUC of 0.516 (SD, 0.004) and 0.488 (SD, 0.009) respectively. With regard to ambient air vs. breath, both approaches yielded AUCs of 1.000 (SD, 0.000) (figures 3(c) and (f)) and confirmed perfect discrimination criterion. These results further support the reliability of our analysis protocol. The AUC of ∼0.5 obtained from predictions of baseline response also suggested that our sample size was large enough to capture sufficient population diversity.

Zoom In Zoom Out Reset image size

For SARS-CoV-2 infection, we found that the pattern-based approach clearly outperformed the molecule-based approach in prediction performance (AUC of 0.849 (SD, 0.004) vs. 0.769 (SD, 0.007)). To illustrate this result, we made use of the subjects' distribution on the PLS coordinate, which allowed us to visualize which approach can better discriminate opposing response classes. We used the complete data set (N = 170) for construction of the PLS coordinate space and plotted subjects' data on the first three PLS components in figures 4(a) (pattern-based) and (b) (molecule-based). The results show significantly better discrimination capability was obtained by the pattern-based approach. The underperformance of the molecule-based approach could potentially be attributed to the exclusion of species with unknown identities in exhaled breath detected by CE-DFCS. As CE-DFCS acquires breath data at extremely high sensitivity, specificity, and dimensionality, applying the pattern-based approach to make full use of the wealth of chemical information collected by CE-DFCS is advantageous in that it bypasses the need for a complete molecular database to directly understand the best possible prediction power.

Zoom In Zoom Out Reset image size

A notable limitation of the pattern-based approach, however, is that it does not reveal which molecules are important for making predictions, but only the optical frequencies at which they are probed. Variable importance analyzed for the pattern-based approach (figure 4(c)) identified prediction-important optical frequencies (VIP scores $\gt$ 1 and highlighted in purple) where measured absorption values were strongly discriminative between SARS-CoV-2 positives and negatives. These frequencies are distributed near-uniformly over the entire spectrum. On the other hand, variable importance analyzed for the molecule-based approach (figure 4(d)) identified a panel of indicative molecular species for SARS-CoV-2 infection: water (H₂O), semiheavy water (HDO), formaldehyde (H₂CO), ammonia (NH₃), methanol (CH₃OH), and nitrogen dioxide (NO₂). Being able to identify the molecules provides better clarity to rationalize a possible prediction. To illustrate, variable importance performed for ambient air vs. breath samples based on the molecule-based approach identified water (H₂O) and semi heavy water (HDO) as the only important predictor variables (data not shown). This is easy to understand because water contents were saturated in breath and hence the machine could solely rely on them for prediction. The panel of indicative molecules identified by the molecule-based approach for SARS-CoV-2 infection provides the opportunity for further studies to elucidate the pathophysiology of SARS-CoV-2 infection.

We analyzed the prediction performance for a list of subject characteristics and potential factors that could confound the results. For prediction of a specific response, subjects from the complete dataset (N = 170) were divided into opposing classes based on the self-reported questionnaire data. Results obtained using the pattern-based approach are presented in figure 5 and the group assignment criteria for different response types are listed in the panels. A summary for all prediction analyses can also be found in table S1 (see supplementary file). From the results, we found random guessing predictions (AUC $\lt$ 0.6) for alcohol use, age, and lactose intolerance, but significant prediction capabilities for smoking, sex, abdominal pain, and constipation (0.6 $\leqslant$ AUC $\lt$ 0.7). On age and abdominal pain, while our subjects had modest correlations with SARS-CoV-2 infection, the significantly better predictive power for SARS-CoV-2 infection suggests that age and abdominal pain do not constitute strong confounders. The superior prediction performance for SARS-CoV-2 infection compared to the list of potential confounders analyzed could potentially be due to SARS-CoV-2 infection eliciting acute and long-term host responses caused by both virus-driven and immune system-associated factors.

Zoom In Zoom Out Reset image size