Feed

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. Exhaled Hydrogen Peroxide for Early Lung Cancer Detection (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 and 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. Analyzing 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 and 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 2132 features, with four key metabolites yielding an AUC of 0.969 for cancer discrimination. 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.