Introduction: Temporomandibular disorder (TMD), oral lichen planus (OLP), and burning mouth syndrome (BMS) are prevalent orofacial conditions whose overlapping symptoms and reliance on subjective clinical assessment complicate differential diagnosis. A rapid, noninvasive, and objective screening tool remains an unmet need.
Methods: Saliva was collected from healthy controls (HC) and patients with TMD, OLP, and BMS (286 samples total) and measured on three-dimensional SERS structures in which M13 bacteriophage serves as a nanoscale gap spacer between gold nanoparticles, generating reproducible hotspots. Spectra were preprocessed (spike removal, Savitzky–Golay smoothing, baseline correction, normalization) and classified using a pathophysiology-informed hierarchical binary MLP framework that first screens disease from HC and then discriminates disease subtypes.
Results and Discussion: The hierarchical framework outperformed a flat multiclass baseline (balanced accuracy 0.684), achieving an AUC of 0.90 for discriminating disease from healthy controls and pairwise inter-disease AUCs up to 0.95, with the BMS and OLP binary classifiers reaching accuracies of 92.1% and 94.7%. Significant inter-group differences at 880, 1230, 1540, and 1605 cm⁻¹—assigned to the tryptophan indole ring mode, Amide III, the tryptophan W3 (C=C) mode, and aromatic residues—link the spectral signatures to disease-associated salivary biochemistry.
Conclusions: Conclusions: Saliva-based SERS on M13 phage-templated structures, combined with hierarchical machine learning, discriminated multiple orofacial diseases noninvasively, achieving an AUC of 0.90 for disease screening and classification accuracies of 71.1%, 92.1%, and 94.7% for the TMD, BMS, and OLP classifiers, respectively. These results support its potential as an adjunctive diagnostic strategy.