Vector Machine Based Decision Tree to Classify Voice Pathologies Using High - Speed Videoendoscopy

Objective/Hypothesis: Little research has explored the potential of computer assisted decision making applied to high-speed videoendoscopy. In this paper, we propose a computer based method for differentiating normal and pathological larynges on the basis of HSV. Methods: HSV recordings were collected from 101 patients with normal larynges, leukoplakia, nodules or polyps. After pre-processing, samples were analyzed for the number of glottal regions present during the open phase, the symmetry of the glottal area, the convex nature of the vocal folds and the ratio of the minimal to maximal glottal area. A decision tree based method with support vector machines at the tree nodes was used to separate samples. Results: Normal samples were differentiated from pathological samples with a sensitivity of 91.1% and a specificity of 81.8%. When samples were divided into normal, nodule, polyp and leukoplakia groups, samples were correctly separated 70.3% of the time. Conclusions: The combination of SVM and decision tree improves the differentiating capabilities of the parameters employed. While our approach was successful in separating normal from abnormal samples, the classification of unique pathologies requires the development stronger individual parameters.