Monitoring human larynx by random forests using questionnaire data

This paper is concerned with noninvasive monitoring of human larynx using subject's questionnaire data. By applying random forests (RF), questionnaire data are categorized into a healthy class and several classes of disorders including: cancerous, noncancerous, diffuse, nodular, paralysis, and an overall pathological class. The most important questionnaire statements are determined using RF variable importance evaluations. To explore multidimensional data, t-Distributed Stochastic Neighbor Embedding (t-SNE) and multidimensional scaling (MDS) are applied to the RF data proximity matrix. When testing the developed tools on a set of data collected from 109 subjects, 100% classification accuracy was obtained on unseen data coming from two - healthy and pathological - classes. The accuracy of 80.7% was achieved when classifying the data into the healthy, cancerous, and noncancerous classes. The t-SNE and MDS mapping techniques facilitate data exploration aimed at identifying subjects belonging to a ”risk group”. It is expected that the developed tools will be of great help in preventive health care in laryngology.