Differentiability of voice disorders through explainable AI

The voice can be affected by various types of pathology. The phoniatric medical examination is the acoustic analysis, which evaluates the characteristic parameters extracted from the vocal signal. Computer-assisted decision-making systems can help specialists to detect vocal pathologies using only the patient's voice. In this study, transfer learning techniques are used to perform the acoustic analysis. Fine-tuned OpenL3 then predicts whether or not the signals contain a pathology by classifying them under 8 different pathologies. A publicly available dataset is used with the categories Hyperkinetic dysphonia, Hypokinetic dysphonia, reflux laryngitis vocal fold nodules, prolapse, glottic insufficiency and vocal fold paralysis in addition to the Healthy class. The results obtained are very convincing. The accuracy with OpenL3, using tranfer learning, was 99.44%. In addition, explainable decision support systems (XDSS) provide an in-depth understanding of the decision-making process. Obtaining an image resulting from the averaging of all the Occlusion Sensitivity maps will enable us to understand the spatio-temporal characteristics of the disordered voices used for classification. Thanks to explainability methods, a new term, the differentiability, can be discussed to explain the black-box operation of deep networks. For purposes of rapid diagnosis and prevention, this work could provide more detail on disordered voices by enabling a promising explainable diagnosis.