Multiplexed Nanomaterial-Based Sensor Array for Detection of COVID-19 in Exhaled Breath.

This article reports on a noninvasive approach in detecting and following-up individuals who are at-risk or have an existing COVID-19 infection, with a potential ability to serve as an epidemic control tool. The proposed method uses a developed breath device composed of a nanomaterial-based hybrid sensor array with multiplexed detection capabilities that can detect disease-specific biomarkers from exhaled breath, thus enabling rapid and accurate diagnosis. An exploratory clinical study with this approach was examined in Wuhan, China, during March 2020. The study cohort included 49 confirmed COVID-19 patients, 58 healthy controls, and 33 non-COVID lung infection controls. When applicable, positive COVID-19 patients were sampled twice: during the active disease and after recovery. Discriminant analysis of the obtained signals from the nanomaterial-based sensors achieved very good test discriminations between the different groups. The training and test set data exhibited respectively 94% and 76% accuracy in differentiating patients from controls as well as 90% and 95% accuracy in differentiating between patients with COVID-19 and patients with other lung infections. While further validation studies are needed, the results may serve as a base for technology that would lead to a reduction in the number of unneeded confirmatory tests and lower the burden on hospitals, while allowing individuals a screening solution that can be performed in PoC facilities. The proposed method can be considered as a platform that could be applied for any other disease infection with proper modifications to the artificial intelligence and would therefore be available to serve as a diagnostic tool in case of a new disease outbreak.

The outbreak of severe acute respiratory syndrome coronavirus-2 (SARS-CoV-2, aka COVID-19) has emerged very rapidly and has already invaded more than 197 countries worldwide, with more than 16 million cases and ∼650 000 deaths within 7 months of the outbreak, as reported by the WHO coronavirus disease situation report 189. About 97.5% of patients develop symptoms of COVID-19 within 11.5 days of exposure, causing late diagnosis and a high infection rate.[1] The molecular tests used so far to confirm COVID-19 are considered the gold standards for SARS-CoV-2 testing. Nevertheless, they require a swab sample and a time-consuming laboratory procedure. Shipping of samples and overload of laboratory facilities entail a delay of many days until the test results are available, increasing the burden to the healthcare system. Epidemiological data based on sequenced viral-RNA show that the spread of COVID-19 has resulted from local community transmission, which means that the source of infection cannot be traced back to a known exposure.[2] Thus, healthcare systems require tests that are rapid, inexpensive, and easy to use for diagnosing or ruling out infection at earlier stages, even before symptoms manifest, to decrease the transmission and mortality rates. Different strategies within the sensors and biosensors solutions have recently been discussed highlighting the importance of point-of-care sensitive systems to cope with the COVID-19 outbreak.[3,4] Indeed some such promising solutions have already been proposed.[5]

Here, we report on a nanomaterial-based sensor array with multiplexed capabilities for detection and monitoring of COVID-19 from exhaled breath. The sensors are composed of different gold nanoparticles linked to organic ligands, creating a diverse sensing layer that can swell or shrink upon exposure to volatile organic compounds (VOCs), causing changes in the electric resistance. In these layers, the inorganic nanomaterials are responsible for the electrical conductivity, with the organic film element providing sites for the adsorption of VOCs.[6,7] When exposed, VOCs diffuse into the sensing layer or fall on the sensing surface and react with the organic segment or the functional groups capping the inorganic nanomaterials. The outcome of the interactions causes a volume change (swelling/shrinkage) in the nanomaterial film.[6] As a result, the contacts among the inorganic nanomaterial block the change (higher/lower) with an increase/decrease of conductivity.[6] The nanomaterial layer exposure to VOCs causes a swift charge transfer to/from the inorganic nanomaterial, producing variations in the measured conductivity even when no steric changes occur within the sensing layer.[6,8]

Different sensors can be used, due to the chemical diversity of the functional group(s) capping the nanoparticles, as an array of cross-reactive semiselective sensory units that mimic the sensing processing method of natural mammalian olfactory systems.[6,9,10] The main advantage of such an array is the flexibility and the possibility that it can be trained to identify a wide range of chemical patterns, “fingerprints”, of varying conditions and for different applications using pattern recognition and machine learning algorithms.[6,9]

The rationale behind this approach relies on findings showing that viral agents and/or their microenvironment emit VOCs[11,12] that can reach the exhaled breath.[8,10,11,13−19] The emergence of VOCs in exhaled breath could occur in early stages of the infection,[9,20] thus serving for immediate detection of COVID-19. The feasibility of this approach as a prescreening diagnostic system was examined via an observational case-control study at the origin of the COVID-19 outbreak, Wuhan, China, during March 2020. Preliminary results and pertinent discussions are presented below.

Results and Discussion

An array consisting of eight gold nanoparticles working on the principle described earlier in the text was developed and integrated with electronic circuitry and an advanced apparatus that collects an exhaled breath sample, by blowing into the device for 2–3 s from a distance of 1–2 cm (see Figure a, Figure S1). A built-in sensor technology advised the study subject when the test was complete. As the breath passes through the array, a mixture of COVID-19-related VOCs reacts with the sensors, and these emit a set of electrical resistance signals as a function of time. An example of a sensor response to different samples can be found in Figure b. If the breath collection was not satisfactory, the subject was asked to repeat the test. Software-based machine learning methods probe the pattern of output signals to get a signature of COVID-19. A more detailed description is provided in the Materials and Methods section, the Supporting Information, and previous work.[21]

Figure 1

(a) Example of breath collection with the developed hand-held breathalyzer system from a patient in Wuhan, China. (b) Representative response of a sensor to three different breath samples. The normalized response of sensor 7 of the breathalyzer system to three different samples: patient A, COVID-19, first sample while infected; patient A, second sample after determined as recovered; and a healthy control. The x-axis represents the cycle measurement; each unit is one cycle of the sensor. The infected sample had a positive change response, while the recovered and control showed negative charges.

Using this nanotechnology, an exploratory clinical study in Wuhan, China, was conducted during March 2020 on 140 participants. The selection of participants from three distinct groups—COVID-19 patients, healthy controls, and non-COVID lung infection/disease controls—is described in Figure .

Figure 2

Patient enrollment and observational design.

The characteristics of the 140 participants are shown in Table S1. Among the patients, more than 60% had no underlying chronic disease, while all the rest suffered at least from hypertension; 10% of those suffered from diabetes mellites. Most patients were nonsmokers (73%), with a mean age of 59 years, with 57% females. For the control group, 67% were nonsmokers, and the mean age was 52 years, with 46% females. For the lung infection control group, 73% were nonsmokers, and the mean age was 63 years, with 44% females.

Statistical analysis was employed on the signals emitted from three binary comparisons: COVID-19 vs control; COVID-19 vs other lung infections; and COVID-19 first vs COVID-19 second sample (Figure ). Seventy percent of the data were used to calculate the discriminant factor analysis (DFA) models. For each comparison, a receiver-operating-characteristic (ROC) curve and cutoff value were determined. The remaining test data were classified blindly on the basis of the cutoff as presented (Figure ). Panels A, B, and C in Figure show data classification from cumulative sensor responses to breath samples as represented by the canonical variable of the discriminant analysis. Panel D shows ROC curves for the cumulative breath-sensor response. The results showed an area under the curve (AUC) of 0.81 [95% CI, 0.70 to 0.89] in patients with COVID-19 vs with controls, 0.97 [95% CI, 0.92 to 0.99] in COVID-19 vs other lung infection/conditions, and 0.87 [95% CI, 0.67 to 1.00] in COVID-19 first sample vs COVID-19 second sample. P < 0.001 was found for the comparisons of the training set for each of the binary classifications.

Figure 3

Diagnosis of COVID-19 patients based on cumulative breath sample response. Panels A, B, and C show data classification from cumulative sensor responses to breath samples as represented by the canonical variable of the discriminant analysis. Box plots of the first canonical score of the training set (70% of samples) and test set (30% of samples). The horizontal dashed line in the box plots represents the cutoff value of the model. True positive (TP), true negative (TN), false positive (FP), false negative (FN). Panel A: COVID-19 patients (n = 41) and healthy controls (n = 57). Panel B: COVID-19 patients (n = 41) and other lung infection/condition controls (n = 32). Panel C: COVID-19 patients at first (n = 41) and second sampling (n = 21). P-values are for the comparisons of the training set for each of two binary classifications. The horizontal line in the boxes represents the median, the cross represents the mean, and the bottom and top of the boxes represent the 25th and 75th percentiles, respectively. Bars represent the upper 90th and lower 10th percentile, and the square dots are outliers. All P-values were adjusted for multiple comparisons using the Tukey–Kramer method. For panel C, the P-value is also adjusted for paired analysis. Panel D shows ROC curves for the cumulative breath-sensor response in patients with COVID-19 (Co) infection compared with controls (C) (black); in COVID-19 infection compared with other lung infection/conditions (LI), (red); and in COVID-19 infection first sample (Co1) compared to COVID-19 infection second sample (Co2) (blue). †P < 0.0001.

Accuracies for the training set varied between 90% and 94% for the three models and between 76% and 95% for the test set (see Table ). The main comparison between COVID-19 and controls gave 100% sensitivity for both training and test groups, while specificity was lower, with 90% in the training set and 61% in the test set; the latter could be due to the safety measurements required for the specific clinical trial or due to the relatively low number of healthy controls as described in the limitations below. The paired analysis of the two sample times showed a clear distinction for the measured breath composition score (P < 0.001; mean difference (95% CI) = −1.57 (−2.27 to −0.87)): in the first sample, all were COVID-19 positive; at the second sample time, all but three were considered cured. Among the three uncured, two were identified by the model as false-negative and one was correctly classified as positive. This could be attributed to the prolonged healing time; it can take a few weeks to reach a definitely cured state.

Table 1

Breath Test Outcomes for the Study Population

	training set			testing setc
statistics	COVID-19 vs controla	COVID-19 vs lung infectionb	COVID-19 1st vs COVID-19 2nd,b	COVID-19 vs control	COVID-19 vs lung infection	COVID-19 1st vs COVID-19 2nd
accuracy (%)	94	90	90	76	95	88
sensitivity (%)	100	90	100	100	100	83
specificity (%)	90	91	69	61	90	100
PPV (%)	88	93	86	61	92	100
NPV (%)	100	87	100	100	100	71
TP (cases)	30	26	32	11	12	10
TN (cases)	35	20	11	11	9	5
FP (cases)	4	2	5	7	1	0
FN (cases)	0	3	0	0	0	2

Classification based on QDA.

Classification based on LDA.

Classification based on the ROC cutoff.

Confounding effects on the main classification were examined; there were no significant differences with respect to age, sex, smoking status, or coexisting conditions (Figure ). The plausibility of the main model built to distinguish COVID-19 from controls to classify all subgroups was tested also. The model showed significant differences between COVID-19 and control between COVID-19 and cured (see SI, Figure S2.).

Figure 4

Evaluation of the confounding factors for COVID-19 patients and control group based on cumulative breath sample response. Panels A, B, C, and D show the data classification of cumulative sensor responses to breath samples as represented by the canonical variable of the discriminant analysis, based on the main DFA model of patients vs controls. Box plots of the first canonical score of the data set of the COVID-19 and control data. Panel A compares male (34 participants) and female (35 participants), P = 0.90. Panel B compares age above 60 years (21 participants) and below 60 years (48 participants), P = 0.11. Panel C compares smokers (21 participants) and nonsmokers (48 participants), P = 0.66. Panel D compares coexisting conditions (18 participants) and noncoexisting conditions (51 participants), P = 0.07. P values are for the comparisons of each set of the confounding factors. None of the P values differ significantly between the confounders, suggesting no effect on the models. The horizontal line in the boxes represents the median, the cross represents the mean, and the bottom and top of the boxes represent the 25th and 75th percentiles, respectively. The I bars represent the upper 90th and the lower 10th percentiles, and the square dots are outliers.

The presented results show excellent sensitivities for the three binary comparisons, while specificities were above average. The examination of the presented nanotechnology tested the assumption that COVID-19 patients can be differentiated not only from controls but also from other lung conditions/infections. The obtained results are comparable to current published data from COVID-19 studies, which suggest 82–98% accuracy for abnormal CT findings and 51–70% sensitivity of the RT-PCR test.[1,22,23] No significant influence of confounding factors was found, similar to previous reports on this technology.[14,19] While specificity levels were relatively low for the main comparison, these values are acceptable for a fast screening or triage process. This is particularly true considering that the current methods have an average sensitivity of 70% and down to 32% in some throat swab cases.[24] The lower specificity means higher false-positives; however in a global perspective, for example, delaying a passenger in the airport for additional 24 h (for a second confirmatory test) is preferable to allowing him/her to board an aircraft and infect 300 passengers. Different prediction models for COVID-19 are built to support medical decision making globally. However, recent reviews indicate that proposed models (mainly diagnostic models based on imaging technology) are either poor or optimistic.[25] Therefore, mass-screening protocols and technologies as presented here can potentially overcome such issues and allow immediate data sharing. We expect that real-time methods such as the exhaled breath approach will significantly reduce unnecessary exposure to contagious persons and support the fight against the COVID-19 pandemic. Moreover, it will reduce the number of unnecessary confirmatory tests and lower the burden on hospitals, while providing individuals with a screening solution that can be performed in PoC facilities.

Without detracting from the success of this breath-based diagnosis of COVID-19, the current study has a few limitations. The case-control design could overestimate sensitivity and specificity, so our results must be interpreted with caution. Also, the results presented apply only to a prediagnosed patient population in China, for which disease duration may vary; further studies for asymptomatic or presymptomatic persons are required in different settings. Owing to safety regulations, part of the controls were sampled in a separate location from those with COVID-19, using a second system, which could raise issues about possible ambient/geographical effects on the VOC measurements. We therefore collected control samples in both locations to identify possible differences. Lastly, breath measured here is composed of respiratory gases, VOCs, and humidity; humidity changes in breath might influence the way sensors work and could have caused the relatively low specificity. In the future, important safety issues with regard to COVID-19 sampling will require special attention to prevent the occurrence of cross contamination within the screening population, either by producing disposable kits or by applying a nonharmful sterilization process when used in public.

Summary and Conclusions

The current article presents a noninvasive intelligent nanomaterial-based hybrid sensor array with multiplexed capabilities for the detection and monitoring of COVID-19-specific VOC mixtures from exhaled breath. The efficacy of this reported nanotechnology was evaluated via a case-control clinical study, with results that are comparable with those recommended by healthcare agencies. Owing to its simplicity and high sensitivity, the reported test could help in rapid screening of large populations and could further be used as a useful rule-out tool for screening of COVID-19 or as part of a triage process. However, a larger cohort study is required to validate the results. The cross-reactive array relies on a training and learning algorithm that will require adaptation and tuning once more data are collected, which will improve classification abilities. Additionally, more confounding factors, mainly other background diseases, will require special attention before such a system can be commercialized. The sensing elements have a limited life span and will require replacement with time. While there is no intention to replace the existing validation molecular tests at this stage, the breath analyzer test proposed here could be potentially designated for rapid large population screening in a short period of time, e.g., at central facilities (airports, shopping centers, train stations, and other public places) or PoC facilities (clinics, hospitals) and as active-case searching in the community, for early detection of the disease in asymptomatic contagious individuals, thus reducing the load on hospitals and preventing other patients from fearing medical care.[26]

Materials and Methods

Study Population

From March 9 to March 27, 2020, a total of 140 participants were enrolled at multiple centers in Wuhan and Hefei, China, as part of an observational study (clinical registry number: ChiCTR2000030556): 49 COVID-19 patients and controls; divided into 58 healthy controls and 33 non-COVID lung infection controls (see Figure ). The study size was calculated for a dichotomous end point; two sample study groups with an alpha of 0.05 and power of 80% with anticipated incidence in the sick group of 99% and 85% in the control result in 47 sick and 94 control samples, respectively. The COVID-19 patients were confirmed by computed tomography (CT), nasal and pharyngeal swab specimens for real-time reverse-transcriptase polymerase-chain-reaction (RT-PCR), and antibody tests (see details in the supplementary clinical data). The enrolled COVID-19 patients were sampled at two time points approximately 3–5 days apart in the COVID-19 intensive care ward assigned to the First Affiliated Hospital of USTC entrusted by the State Council of China located in Wuhan Union Hospital, China. The research protocol was approved by the ethics committee of Anhui Provincial Cancer Hospital, West District of The First Affiliated Hospital of USTC. All participants provided written informed consent.

Nanomaterial-Based Hybrid Sensor Array

The developed portable hand-held analyzer comprises eight different sensors (see Figure S1). The nanomaterial-based sensor array that is used to analyze the breath samples contains cross-reactive, chemically diverse chemiresistors based on stabilized spherical gold nanoparticles (GNPs; core diameter: 3–4 nm) with various and diverse organic ligands as previously described.[21] The electrodes of the chemiresistors were fabricated on silicon wafers in clean room facilities (Technion, Israel) by only two photolithography steps. These were metal evaporation and lift-off. Silicon wafers capped with a thermal silicon oxide film of 1 μm were used (purchased from Nova Electronic Materials, LLC, USA). The circular interdigitated platinum electrodes were deposited by an electron-beam evaporator (Evatec BAK501). The outer diameter of the circular electrode area was 1000 μm; the gap between the two adjacent electrodes and the width of each electrode were both 10 μm. The printed sensors were produced by inkjet printing using a dispenser system, sciFLEXARRAYER S3 (SCIENION AG, Berlin, Germany). The dispenser’s deposition process is by a noncontact piezo-dispensing system that allows precise drops of 0.3 ± 0.05 nL to be dispersed. The sensors were fabricated by dropping the GNP solution exactly on the middle of the electrodes. After the GNP synthesis had been analyzed, each solution was printed on the eight electrodes using the automatic dispenser system.

Breath Sampling

Breath samples were collected by the study subjects breathing directly into the aperture of the instrument for at least four seconds, keeping the instrument approximately 1–2 cm from the mouth. The overall composition of the breath samples collected includes normal expiratory gases together with the different VOCs and humidity. Built-in sensor technology advised the study subject when the test was complete. If the breath collection was not satisfactory, the subject was asked to repeat the test. Detailed information on the procedure is provided in the supplementary methods. The primary outcome was determination of the change of VOC metabolites in the breath, to assess the plausibility to differentiate individuals with COVID-19 from healthy controls. The secondary outcome was to differentiate individuals with COVID-19 from other non-COVID-19 conditions.

Feature Extraction and Statistical Analysis

Features were extracted from the output files of the breath samples for all sensors in the array. An example of a sensor response to different samples can be found in Figure . The feature was calculated as the change in electrical resistance between the breath signal and the baseline signal divided by the baseline signal. The tested groups were subjected to binary comparison, and the data were divided randomly into training sets (70% samples) and test sets (30% samples). The results of DFA on the training set were validated using the test set. The base analysis between the COVID-19 and control samples used the quadratic DFA model based on the measurements of three sensors. For the other two subcomparisons (COVID-19 vs other lung infections; COVID-19 first vs second sampling), the same three sensors as for the main model were used by applying two binary linear DFA models. The model performance of the training set was first determined by measuring the AUC of the ROC and was used to calculate the cutoff values on the basis of Youden’s index, which classifies the tested groups as giving either a positive or a negative result for the test set classification. Subsequently, other parameters of model performance were analyzed including accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). To address the influence of the main confounding factors on breath analysis, age, sex, and smoking status were registered and evaluated on the basis of the same model classifier achieved by DFA. The first and second samplings from the same COVID-19 patients (i.e., correlated variables) were compared using paired t test analysis (by subject) only between patients who had provided samples at the two time points. The matched platform results with a Tukey mean difference presented the mean difference and 95% confidence interval (CI). All P-values were adjusted for multiple comparisons using the Tukey–Kramer method. Significant differences in the one-way and/or matched pair test were considered at a cutoff P-value of 0.05 between the subgroups checked, as determined from the results using JMP Pro, version 14.0.0 (SAS Institute Inc., Cary, NC, USA, 1989–2005). Of the 140 samples, 10 were excluded after collection for technical reasons, either bad sampling or sensor failure during sampling and before statistical analysis.