Point-of-care breath sample analysis by semiconductor-based E-Nose technology discriminates non-infected subjects from SARS-CoV-2 pneumonia patients: a multi-analyst experiment

Abstract

Metal oxide sensor-based electronic nose (E-Nose) technology provides an easy to use method for breath analysis by detection of volatile organic compound (VOC)-induced changes of electrical conductivity. Resulting signal patterns are then analyzed by machine learning (ML) algorithms. This study aimed to establish breath analysis by E-Nose technology as a diagnostic tool for severe acute respiratory syndrome coronavirus type 2 (SARS-CoV-2) pneumonia within a multi-analyst experiment. Breath samples of 126 subjects with (n = 63) or without SARS-CoV-2 pneumonia (n = 63) were collected using the ReCIVA® Breath Sampler, enriched and stored on Tenax sorption tubes, and analyzed using an E-Nose unit with 10 sensors. ML approaches were applied by three independent data analyst teams and included a wide range of classifiers, hyperparameters, training modes, and subsets of training data. Within the multi-analyst experiment, all teams successfully classified individuals as infected or uninfected with an averaged area under the curve (AUC) larger than 90% and misclassification error lower than 19%, and identified the same sensor as most relevant to classification success. This new method using VOC enrichment and E-Nose analysis combined with ML can yield results similar to polymerase chain reaction (PCR) detection and superior to point-of-care (POC) antigen testing. Reducing the sensor set to the most relevant sensor may prove interesting for developing targeted POC testing.

Keywords: COVID‐19; E‐Nose; breath gas; machine learning; mass spectrometry; metal oxide sensor; pneumonia; volatile organic compounds.

FIGURE 1

Team A, experiment I: ROC curves and feature importance. ROC curves depicting the sensitivity (y‐axis) and 1 − specificity (x‐axis) for the random forest (RF) (A, AUC: 0.94) and Glmnet classifier (B, AUC: 0.90). The bold line shows the micro‐averaged ROC curve based on the resampling results with its pointwise confidence interval. Panel (C) depicts variable importance (mean decrease in impurity) for the top 40 features of the RF classifier and panel (D) shows regression coefficients for the top 40 features of the Glmnet classifier.

FIGURE 2

Team A, experiment II: ROC curves and feature importance. Performance of random forest (RF) (A, AUC: 0.91) and Glmnet (B, AUC: 0.92) classifier on data generated via feature extraction. The micro‐averaged ROC curve is depicted (bold line) with the corresponding confidence interval (gray). The impact of different generated features on classification accuracy is shown through variable importance for the RF classifier (C), and regression coefficients for the Glmnet classifier (D).

FIGURE 3

Team B: ROC curves and importance maps for experiments III and IV. ROC curves (A and B) represent the performance of the learner on data from all sensors. Comparing classification impact of features on importance maps shows that sensor 9 has the highest importance for classification (C and D). While experiment III (A and C) obtained results without hyperparameter tuning, experiment IV (B and D) depicts performance of the model with hyperparameter tuning.

FIGURE 4

Team B: ROC curves and importance map for experiments V and VI. ROC curves show results when the learner was only provided with data from sensor 9 without (A) or with (B) hyperparameter tuning. As for experiments III and IV, the highest impact can be observed during the initial phase of measurements (C, without hyperparameter tuning; D, with hyperparameter tuning), corresponding to peak signals of raw data (see Figure 6).

FIGURE 5

Team C, experiment VII: precision P(C|+) and negative predictive value P(nC|‒) versus prevalence P(C). (A) Performance of random forest (RF) learner on data of sensor 9 from lower respiratory tract samples is dependent on the prevalence of severe acute respiratory syndrome coronavirus type 2 (SARS‐CoV‐2) in the test population. (B and C) Performance of the model is also dependent on the amount of training data provided to the learner during training. Depicted are precision (B) and negative predictive value (C) versus prevalence from sensor 9 data obtained from lower respiratory tract samples.

FIGURE 6

Exemplary signal pattern generated by E‐Nose analysis. Curves depict changes in conductivity of 10 different metal oxide semiconductor sensors over time. Each line indicates one sensor. Conductivity (G) is normalized for each sensor's zero value (G0) measured before injection.

FIGURE 7

Study design and multi‐analyst approach. Schematic depiction of the study workflow and comparison of different analytical approaches and foci of data processing. PCR, polymerase chain reaction; EDU, Enrichment and Desorption Unit; SARS‐CoV‐2, severe acute respiratory syndrome coronavirus type 2.