Voice disorder detection using machine learning algorithms: An application in speech and language pathology

The healthcare industry is currently seeing a significant rise in the use of mobile devices. These devices not only provide ways for communication and sharing of multimedia information, such as clinical notes and medical records, but also offer new possibilities for people to detect, monitor, and manage their health from anywhere at any time. Digital health technologies have the potential to improve patient care by making it more efficient, effective, and cost-effective. Utilizing digital devices and technologies can have a positive impact on many health conditions. This research focuses on dysphonia, a change in the sound of the voice that affects around one-third of individuals at some point in their lives. Voice disorders are becoming more common, despite being often overlooked. Mobile healthcare systems can provide quick and efficient assistance for detecting voice disorders. To make these systems reliable and accurate, it is important to develop an algorithm that can classify intelligently healthy and pathological voices. To achieve this task, we utilized a combination of several datasets such as Saarbruecken voice dataset (SVD), the Massachusetts Eye and Ear Infirmary database (MEEI), and a few private datasets of various voices (healthy and pathological) Additionally, we applied multiple machine learning algorithms, including decision tree, random forest, and support vector machine, to evaluate and determine the most effective algorithm among them for the detection of voice disorders. The experimental analyses are performed in terms of sensitivity, accuracy, receiver operating characteristic area, specificity, F-score and recall. The results demonstrated that the support vector machine algorithm, depending on the features selected by using appropriate feature selection methods, proved to be the most accurate in detecting voice diseases.