A Robust KNN Approach for Multi-Class Laryngeal Disease Detection using MFCC Features

Early detection of laryngeal diseases is vital for preventing progression and improving patient outcomes. Speech signals offer a convenient, non-invasive biomarker, and Mel-frequency cepstral coefficients (MFCCs) effectively capture their acoustic properties. This study introduces a graph-based KNN framework that constructs neighborhood graphs from MFCC sequences and employs Graph Neural Networks (GNNs) to model local topological relationships. Evaluations on a dataset of 320 subjects across five categories (healthy, laryngeal cancer, polyps, nodules, and leukoplakia) show that the proposed method consistently outperforms baseline models, achieving 0.96 accuracy in binary classification and 0.88 in five-class classification. These results highlight the robustness of graph-based modeling and its potential for non-invasive, voice-based screening of laryngeal diseases.