Answers are grounded in this paper and cite their sources.
Background: Lung cancer and pneumonia can present with non-specific symptoms that overlap with other respiratory conditions. This study evaluated whether breath measurements from the Inflammacheck® device contain sufficient diagnostic information to support machine-learning classification of lung cancer and pneumonia. Methods: A secondary analysis of 282 participants with asthma, chronic obstructive pulmonary disease, interstitial lung disease, bronchiectasis, breathing pattern disorder, pneumonia, lung cancer and healthy controls was undertaken. Seven device-derived variables comprising H₂O₂, exhalation duration, sample collection time and up to 6-minutes of tidal-breathing based peak CO₂, breath humidity, breath temperature, exhalation flow rate were analysed. Eight-class and three-class shrinkage linear discriminant analysis (LDA) models were evaluated using stratified 10-fold cross-validation with out-of-fold predictions. Class-specific probability thresholds were scanned within prespecified 85–95% specificity ranges. An exploratory soft-voting ensemble combining LDA, Extra Trees, radial-basis-function support-vector machine and XGBoost was also evaluated. Results: Eight-class classification showed limited discrimination, whereas consolidation into three classes (lung cancer, pneumonia and other respiratory disease) improved diagnostic separation. At operating thresholds corresponding to approximately 89–90% specificity, the 3-class LDA demonstrated discriminatory ability for both lung cancer and pneumonia, with AUROCs of 0.650 and 0.728, respectively. The soft-voting ensemble model further improved discrimination, increasing AUROC to 0.725 for lung cancer and 0.753 for pneumonia. Conclusions: Inflammacheck® variables contain promising measurable discriminatory information for further development of a screening-tool. Future modelling of the underlying continuous breath time-series (“breathprint”) using deep-learning approaches may extract additional discriminatory features not captured by participant-level summary measures.
No takes yet. Share an insight, caveat, or question.
D'Cruz et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: