Early detection of lung cancer remains a clinical priority to improve outcomes and guide timely referral. The ExPeL study evaluated a noninvasive, point-of-care breath test approach using the Inflammacheck device, which measures exhaled breath condensate (EBC) hydrogen peroxide and several physiological parameters. The approach combines these measurements with supervised machine learning to discriminate lung cancer cases from controls in a real-world screening population drawn from the UK Targeted Lung Health Check (TLHC) programme.
Participants for ExPeL were recruited from the TLHC screening programme and comprised individuals with suspected lung cancer (cases) and low-risk ever-smoker controls. The study analyzed participants who provided valid EBC data via Inflammacheck. In the reported cohort, 34 participants had valid EBC measurements; among cancer cases, 83% were early-stage (stage I–II), consistent with a screening population. The abstract does not report the total number initially recruited, detailed inclusion/exclusion criteria, demographic breakdown, or exact counts of cases versus controls beyond the valid-EBC subset.
The Inflammacheck device collected exhaled breath condensate and recorded multiple variables: EBC H2O2 concentration, end-tidal CO2, humidity, temperature, and exhalation flow rate. These outputs were treated as intrinsic physiological and biochemical markers derived at point of care. The device-derived variables were used for multivariate and supervised machine learning analyses to assess group separation and develop classifiers.
Multivariate analyses applied to the EBC-derived variables included principal component analysis (PCA), linear discriminant analysis (LDA), and mapping with Mahalanobis distance. These analyses demonstrated clear separation between lung cancer cases and controls, with greater dispersion observed among cancer patients, suggesting underlying physiological heterogeneity that may not be apparent from univariate measures.
The study used a SMOTE-balanced dataset to address class imbalance during model training. Supervised models were constructed using ensemble approaches, including stacked and voting ensembles. Models were trained on the balanced data and evaluated on held-out test sets. The reported top-performing model, the voting ensemble, produced the following performance metrics on the test set: Accuracy 85.7%, Sensitivity 80%, Specificity 100%, Precision (positive predictive value) 100%, ROC-AUC 0.90, and Matthews correlation coefficient (MCC) 0.73. The authors emphasize that no false positives were identified in the evaluated test data. The abstract does not report confidence intervals, absolute counts of true/false positives and negatives, or external validation in independent cohorts.
In parallel to device-based physiological measurements, the study conducted an untargeted liquid chromatography–mass spectrometry (LC–MS) metabolomics screen on EBC samples. This untargeted analysis detected 2,132 molecular features. From those features, four metabolites were identified as key discriminatory markers; a combined model using these four metabolites achieved an area under the receiver operating characteristic curve (AUC) of 0.969 for distinguishing cancer from controls. The abstract does not list the metabolite identities or provide details of statistical selection, validation, or potential overlap with device-derived classifier results.
The authors report that Inflammacheck, combining point-of-care physiological and biochemical breath measures with machine learning, can effectively distinguish early-stage lung cancer in a screening-relevant cohort. The high specificity and absence of false positives in the reported test set are highlighted as advantageous for primary care and screening triage, where a high false-positive rate can generate unnecessary investigations. The metabolomics results suggest complementary molecular signatures in EBC that may further enhance discrimination.
The abstract does not provide detailed methodological information such as the full sample size recruited, demographic and clinical characteristics of cases and controls, statistical confidence intervals, or external validation in independent populations. Information on the reproducibility of device measurements across settings, feasibility in routine primary care, and prospective impact on referral pathways was not reported in the abstract. Conflict-of-interest disclosures noted that several authors reported no conflicts, one author reported payments and advisory roles with multiple pharmaceutical companies, and one author serves as a Chief Respiratory Officer on a hospital NHS Trust board.
In this screening-derived cohort, the ExPeL study found that the Inflammacheck breath test, using EBC H2O2 and physiological parameters with machine learning, discriminated early-stage lung cancer from controls with high reported specificity and strong overall performance (ROC-AUC 0.90). Complementary untargeted LC–MS metabolomics identified molecular features with an AUC of 0.969 when combined. The abstract supports further evaluation but does not replace the need for larger, prospective validation and reporting of detailed methods and external validation before clinical implementation.