Artificial Neural Network-based Classification to Screen for Dysphonia Using Psychoacoustic Scaling of Acoustic Voice Features

For diagnosis and classification of dysphonia, voice specialists can choose from an array of diagnostic tools like perceptual tests or acoustic voice analysis. These methods have in common that they require a high level of specialized training and experience, and therefore are mostly reserved to specialized centers. We aimed at developing an acoustic voice analysis system that could be used as a screening device to monitor, document, and diagnose voice problems that are also encountered by non-voice specialists, such as anesthesiologists, head and neck surgeons, and general surgeons before surgery of the thyroid gland and the upper thoracic aperture. An acoustical feature extraction paradigm that focused on jitter, shimmer, standard deviation of fundamental frequency, and the glottal-to-noise excitation ratio was used to reanalyse 120 voice samples previously analyzed by Schönweiler et al (A Novel Approach to Acoustical Voice Analysis Using Artificial Neural Networks. JARO. 2000:1;270-282). An improved artificial neural network (ANN) was used for classification. Building on this preliminary work, we modified the mathematical algorithm to further improve classification accuracy. Eighty percent of all voice samples could be classified correctly as either healthy or hoarse (sensitivity: 63.0%; specificity: 93.9%; area under the curve: 0.854). The adaptation of the ANN-voice analysis system for mobile use may facilitate its use and acceptance by non-voice specialists for the discovery and documentation of preexisting voice disorders, and may thereby lead to a timely initiation of further diagnosis and therapy by voice specialists.