Explainable Artificial Intelligence-Driven Salivary Exosome Spectroscopic Profiling for Clinical Diagnosis and Metastasis Detection of Oral Cancer

Oral squamous cell carcinoma (OSCC), the most common oral malignancy, requires accurate diagnostic methods for patient stratification and treatment guidance. Exosome-based liquid biopsy represents a promising minimally invasive approach to cancer detection. This study develops an explainable artificial intelligence (xAI)-assisted label-free surface-enhanced Raman spectroscopy (SERS) platform for profiling salivary exosomes to enable noninvasive OSCC diagnosis and metastatic stratification. A fully connected artificial neural network is designed to extract discriminative features from complex SERS data. Trained on cellular exosome SERS data sets, the model achieves 90.63% accuracy in distinguishing OSCC patients from healthy subjects and 86.63% accuracy in differentiating nonmetastatic and metastatic OSCC cases. Importantly, Shapley additive explanation-based xAI interpretation identifies tryptophan residues in transmembrane proteins as regulators of carcinogenesis, while genetic mutations are linked to metastatic progression, thereby bridging diagnostic outcomes with molecular mechanisms. This work establishes a biochemically interpretable SERS-xAI framework for cancer diagnosis, advancing precision oncology through mechanistic insights.