Machine Learning-Enhanced Fiber-Optic Raman Spectroscopy for In Situ Oral Cancer Classification

A portable Raman spectroscopy system was developed for in vivo diagnosis of oral squamous cell carcinoma (OSCC) in clinical settings. 46 patients with histopathologically confirmed OSCC were involved. The preprocessed dataset was used to train four supervised machine learning classifiers, support vector machine (SVM), k-nearest neighbors (KNN), ensemble algorithms (EA), and linear discriminant analysis (LDA), under a stratified ten-fold cross-validation scheme. Comparative evaluation demonstrated that SVM achieved perfect classification accuracy (100%), outperforming LDA (96.6%), EA (90.7%), and KNN (83.3%). SVM performance was evaluated against retained PC count to assess robustness and generalizability. The area under the receiver operating characteristic curve (AUC-ROC) remained above 80% with two PCs and increased to 96% when eight PCs were retained. These findings underscore the clinical translational potential of machine learning driven Raman spectroscopy, with high discriminative power and generalizability for oral cancer.