Towards noninvasive screening for malignant tumours in human larynx.

This article is concerned with soft computing-based noninvasive screening for malignant disorders in human larynx. The suitability of two types of data for the analysis is explored. The questionnaire data and the digital voice recordings of the sustained phonation of the vowel sound /a/ are the data types considered in this study. The screening is considered as a task of data classification into the healthy, cancerous, and noncancerous classes. To explore data and decisions a nonlinear mapping technique exhibiting the property of local data ordering is applied. The classification accuracy of over 92% was obtained for unseen questionnaire data collected from 240 subjects. The experimental investigations have shown that, concerning the three classes, the questionnaire data carry much more discriminative information than the voice signal. Two-dimensional plots created using the mapping technique provide further insights into the data and decisions obtained from the classifiers.