Machine learning-assisted chemometric analysis of salivary ATR-FTIR spectra for the diagnosis of oral submucous fibrosis

Background Oral submucous fibrosis (OSMF) is a chronic, potentially malignant disorder strongly linked to areca nut chewing. Clinical diagnosis is challenging, and histopathology, though the gold standard, is invasive and subject to inter-observer variability. Saliva analysis using Fourier Transform Infrared (FTIR) spectroscopy offers a promising non-invasive alternative. This study evaluated salivary ATR-FTIR spectral analysis integrated with machine learning for classification of OSMF, individuals with adverse oral habits, and healthy controls. Methodology Unstimulated saliva samples were collected from 450 participants comprising 150 patients with clinically diagnosed OSMF, 150 individuals with adverse oral habits without clinical evidence of OSMF, and 150 healthy controls. After centrifugation, the saliva supernatant was analyzed by ATR-FTIR spectroscopy over 400–4000 cm − ¹. Spectra underwent baseline correction and normalization followed by Savitzky–Golay smoothing. For machine-learning analysis, Standard Scaler and PCA were incorporated within the classification pipeline. Decision Tree, Random Forest, support vector machine (SVM), and XGBoost classifiers were evaluated using stratified 10-fold cross-validation. Hyperparameter optimization and performance estimation used the same cross-validation framework. An independent cohort of 90 participants was used for validation. Results XGBoost showed the highest cross-validation performance, with accuracies of 90.62% for Normal versus Habit, 97.67% for Normal versus OSMF, and 91.30% for three-class classification. Statistical analysis identified 157 nominally significant wavenumbers (p<0.05) within a continuous region from 1581 to 2142 cm − ¹ corresponding to protein and collagen bands. Principal component analysis showed that PC1 and PC2 explained 89.26% and 6.95% of the total variance, respectively (96.21% cumulatively). In the independent validation cohort, accuracy was 80.0% for OSMF versus Normal and 71.7% for Habit versus Normal. Conclusion Salivary ATR-FTIR combined with machine learning demonstrated promising classification performance and preliminary performance in an independent cohort. The identified spectral region is biologically plausible in the context of altered salivary biochemical composition associated with inflammation and fibrosis, but it should not be regarded as a validated disease-specific biomarker. Salivary ATR-FTIR is best considered a complementary, non-invasive screening and risk-stratification approach rather than a replacement for clinical/oral examination or histopathological assessment when biopsy is indicated. Larger, diverse, multicenter studies with rigorous validation are needed before clinical implementation.