MACHINE LEARNING IN SPECTROSCOPIC ANALYSIS OF SALIVA: EVALUATION OF THE EFFECTIVENESS OF CLUSTERING AND DEEP NEURAL NETWORKS FOR THE DETECTION OF ORAL CANCER

Oral cancer (OC) remains a serious oncological problem, as late detection significantly reduces patient survival. Raman spectroscopy is a promising method for non-invasive screening; however, its effectiveness depends on the chosen analytical approach. Unsupervised clustering methods often fail to identify clear groups in spectral data, limiting their applicability. In contrast, convolutional neural networks (CNNs) can detect hidden patterns in spectral data. Objective – to evaluate the effectiveness of clustering methods and CNNs in the saliva spectroscopic analysis for OC screening and determine the optimal approach for spectral classification. Materials and Methods. Spectral data from four patient groups were analyzed: healthy individuals, patients with precancerous conditions (leukoplakia, lichen planus), and patients with oral cancer. Various clustering methods (PCA, t- SNE, K-Means, DBSCAN, GMM, autoencoder + K-Means) and CNNs were applied. Results. Clustering methods failed to form distinct groups, as indicated by low ARI (0.02-0.05) and silhouette coefficient (~0.1). CNNs achieved an accuracy of 84.6% and an AUC of 0.86 when using the full spectral range. Conclusions. Unsupervised clustering methods proved ineffective due to the high overlap of spectra from different groups. In contrast, CNNs demonstrated high classification accuracy, making them a promising tool for automated OC screening.