Abstract
The rapid diffusion of voice disorders and the lack of appropriate knowledge about the problem have prompted the search for novel and reliable approaches to detect dysphonia, through easy and accessible instruments such as mobile devices. These systems represent, in fact, valid instruments to improve the patient care not only to facilitate the monitoring of symptoms of any diseases but also supporting the correct diagnosis of pathology, such as the dysphonia. In this paper, we propose a new marker, namely the dysphonia detection index, able to support the evaluation of voice disorders, which can be embedded in a mobile health solution. Four acoustic parameters are combined in a single marker to globally evaluate the state of the health of the voice and to assess the presence or not of a voice disorder. A model tree regression algorithm has been applied to define the relationship between these parameters, and the Youden analysis has been used to define the threshold value to distinguish a pathological from a healthy voice. The reliability of the proposed index has been tested in terms of correct classification of accuracy, sensitivity, and specificity. A dataset of 2003 voices has been used to evaluate the performance of our proposed index, composed of samples selected from three different databases: the Massachusetts Eye and Ear Infirmary, the Saarbruecken Voice, and the VOice ICar fEDerico II databases. Our approach achieved the best performances in comparison with other algorithms, and accuracy equals to 82.2%, while sensitivity and specificity are 82% and 82.6%, respectively.