Integration of Electronic Nose Profiles from Breath, Urine, and Axillary Sweat for Colorectal Cancer Classification

Electronic noses provide a pattern-based approach for investigating volatile profiles associated with colorectal cancer; however, the combined analysis of participant-matched sensor profiles from different biological matrices remains largely unexplored. This study evaluated breath, urine, and axillary sweat profiles acquired from the same participants using a single microelectromechanical system (MEMS)-based electronic nose system for colorectal cancer (CRC) classification. Matched records from 60 participants were analyzed, including 27 with CRC and 33 controls (CO). Signals from each biological matrix were processed independently using fractional-difference feature extraction, min–max scaling, Orthogonal Signal Correction (OSC), and Principal Component Analysis (PCA). Because OSC is a supervised preprocessing method that incorporates class information, the PCA score plots were used only for descriptive visualization of the transformed feature space, whereas predictive performance was assessed exclusively within the cross-validation framework. The selected PCA scores were subsequently integrated into a six-feature participant-level representation. Five supervised classifiers were evaluated using stratified five-fold cross-validation, with all data-dependent transformations fitted exclusively on the corresponding training sets. In this exploratory setting, Logistic Regression showed the highest observed mean accuracy of 98.3 ± 3.3% among the evaluated models, though this represents an internal estimate rather than a claim of statistical superiority. Aggregated out-of-fold predictions yielded 98.3% accuracy, 100.0% sensitivity, 97.0% specificity, a 98.2% F1-score, and an AUC of 1.00 (95% CI: 0.99–1.00). Our comparative analysis revealed that axillary sweat alone demonstrated a high discriminative power and that the pairwise combination of urine and sweat achieved 100% accuracy. While the complete three-matrix integration maintained high robustness, it did not provide incremental value over sweat alone. These findings highlight the remarkable potential of the axillary volatilome for CRC classification and suggest that selectively combining highly discriminative matrices may be more effective than integrating all available biological sources.