Harnessing exhaled breath for lung cancer early detection – results from the ExPeL study

ABSTRACT Background Scalable, non-invasive tools are critically needed to improve early lung cancer detection and optimize primary care referral pathways. We evaluated Inflammacheck®, a point-of-care device utilizing exhaled breath condensate (EBC) H₂O₂ and physiological parameters with machine learning, for non-invasive lung cancer detection in a real-world screening population. Methods ExPeL study participants, from the UK Targeted Lung Health Check (TLHC) programme, included individuals with suspected lung cancer and low-risk ever-smoker controls. EBC was collected via Inflammacheck®, measuring H₂O₂, end-tidal CO₂, humidity, temperature, and exhalation flow rate. Multivariate analyses (PCA, LDA, Mahalanobis distance) assessed intrinsic group separation. SMOTE-balanced data trained supervised machine learning models (stacked and voting ensembles), which were then evaluated on held-out test sets. In parallel, untargeted LC–MS metabolomics was performed to identify discriminatory molecular features. Results Analysing 34 participants with valid EBC data, 83% of cancer cases were early-stage (I–II), reflecting a screening population. Multivariate analysis clearly separated lung cancer and controls across PCA, LDA, and Mahalanobis mapping. The voting ensemble model achieved: Accuracy 85.7%, Sensitivity 80%, Specificity 100%, Precision (PPV) 100%, ROC–AUC 0.90, MCC 0.73. Crucially, no false positives were identified. EBC variables revealed greater dispersion in cancer patients, reflecting physiological heterogeneity missed by univariate analysis. Untargeted metabolomics identified 2,132 features, with four key metabolites yielding an AUC of 0.969 for cancer discrimination. Discussion Inflammacheck® effectively distinguishes early-stage lung cancer via a rapid, non-invasive breath test, findings which are highly relevant for primary care and screening triage, where non-specific symptoms and low prevalence pose challenges.