Explainable AI-based breath metabolite profiling for early lung cancer detection.

Lung cancer (LC) is a major cause of cancer-related mortality and current screening methods are invasive, costly or radiation-intensive. This study aims to develop a machine-learning-framework using electronic nose (E-nose) sensor data and demographic variables to classify LC in a cohort of 118 participants (65 LC, 53 healthy controls). Breath samples were analyzed using a six-sensor metal-oxide semiconductor array, with each sensor operated at three different temperatures, resulting in 18 sensor-response features. Regularized logistic regression, random forest and XGBoost classifiers were trained and evaluated using nested five-fold cross-validation with three-fold inner hyperparameter tuning, on sensor-only, demographic-only and combined feature sets with and without SMOTE class balancing. Sensor-only models achieved the highest performance across all three classifiers (ROC-AUC up to 0.9611 ± 0.0357), while demographic-only models performed substantially lower (ROC-AUC 0.7566 ± 0.0468- 0.7815 ± 0.0833). The addition of demographic features did not provide a performance improvement over the sensor-only model. Out-of-fold SHAP analysis identified sensor S4, across all three temperature regimes as the most influential predictor, consistent with univariate statistical testing. A nested feature-ablation analysis showed that a stable subset of ten SHAP-ranked features retained performance comparable to the full feature set (ROC-AUC of 0.9575 ± 0.0462 vs 0.9529 ± 0.0522). Age and smoking status were significantly associated with disease status but contributed limited additional predictive value beyond sensor measurements. These findings support sensor-derived breath VOC patterns as a promising interpretable, non-invasive signal for LC screening, supporting further evaluation of E-nose based screening in large, multi-centre cohorts.