Identification of oral squamous cell carcinoma by an electronic nose based on an array of metal oxide sensors with machine learning.

BACKGROUND

Oral squamous cell carcinoma (OSCC) presents a significant diagnostic challenge due to the lack of specific early-stage symptoms, in which the rapid, non-invasive tools are urgently needed. However, the utility of electronic nose (eNose) technology as an adjunct diagnostic aid remains to be validated under standardized sampling conditions.

METHODS

We performed a diagnostic study to evaluate whether a portable "PEN3" equipped with 10 metal oxide sensors (eNose) could identify OSCC patients. We analyzed aspirated air samples from OSCC patients (n = 98) and healthy controls (n = 76) using the eNose, with histopathological diagnosis serving as the reference standard. To analyze eNose signals, we adopted a dual-branch modeling strategy: five data-driven machine learning classifiers enhanced by Kernel Principal Component Analysis (KPCA), and a Weighted Least Absolute Shrinkage and Selection Operator (Lasso) model designed to incorporate clinical prior knowledge.

RESULTS

Analysis of 174 breath samples revealed that standard machine learning models (specifically SVM) could distinguish OSCC with an AUC exceeding 94%. Crucially, the Physiologically-Weighted Lasso model achieved comparable robust performance (AUC = 91.71%, Sensitivity = 90.13%) without relying on complex non-linear manifolds. We found that Physiologically-Weighted Lasso model achieved more stabilized result comparing the unweighted model. Moreover, SHapley Additive exPlanations (SHAP) confirmed that the model (alcohols, carbonyls, organic sulfides, and alkanes) effectively prioritized sensors responsive to organic sulfides.

CONCLUSIONS

We developed a promising eNose-based diagnostic model that effectively balances high diagnostic accuracy with clinical interpretability. The validation of the Physiologically-Weighted Lasso model indicates distinct gas metabolite patterns for OSCC, suggesting that the device detects a genuine pathological metabolic shift driven by organic volatiles rather than stochastic noise. This provides a transparent and potential adjunct tool for non-invasive OSCC screening. TRIAL REGISTRATION: Chinese Clinical Trial Registry, ChiCTR2500102625. Registered 16 May 2025—Retrospectively registered.