Real-time breath metabolomic profiling may detect lung cancer–associated breath features, but exploratory case–control findings require cautious interpretation because small cohorts are susceptible to overfitting and may not reflect real-world diagnostic performance.

In this single-centre matched case–control study, adults undergoing bronchoscopy for suspected lung cancer provided real-time whole-breath samples before bronchoscopy. In confirmed malignancy, bronchoscopic airway samples were collected under general anaesthesia to evaluate whole-breath concordance. Controls were matched by age, sex and smoking history. Machine-learning analyses used nested participant-level cross-validation.

Seventy-six participants were included: 38 in the malignancy cohort, comprising 37 primary lung cancers and one thymic sarcomatoid carcinoma, and 38 matched controls. Most cancers were stage III–IV. Feature selection identified 23 breath VOC features contributing to separation between lung cancer and controls. Twenty-two features were also detected in paired airway samples; feature-level correlations provided partial whole-breath–airway concordance, and formal equivalence testing showed no systematic ipsilateral–contralateral airway differences. Stage-related analyses were exploratory because only 13 malignancy cases had stage I–II disease. Within this enriched cohort, participant-level out-of-fold analysis showed exploratory signal separation using the primary 23-feature Elastic Net model (AUC 0.865, 95% bootstrap CI 0.769–0.938), but this estimate should not be interpreted as real-world diagnostic accuracy.

Real-time breath profiling identified a lung cancer–associated breath feature pattern in this exploratory matched case–control cohort. The findings support prospective multicentre validation with locked algorithms before clinical diagnostic performance can be inferred in lower-prevalence indeterminate nodule and screening populations.

NCT02781857 (25/May/2016).