Neurological Patient Voice Classification With Deep Learning Techniques

Voice analysis is a potential tool that can be used as an objective parameter in the diagnosis of neurological diseases. The disease process can affect the voice signals of individuals with neurological diseases to different degrees. This increases the usability of voice in disease diagnosis. This study aims to develop a method based on voice analysis in the diagnosis of neurological diseases. The study presents a voice-based classification method for individuals with neurological diseases. For this purpose, individuals diagnosed with neurological diseases were asked to say a sentence and voice recordings were made. The audio recordings obtained were processed in a computer environment and used to classify whether the person was a Multiple Sclerosis (MS) patient, a healthy individual, or had one of the other neurological diseases. The data set was created by converting audio recordings into visual representations called spectrograms using signal processing techniques such as Fast Fourier Transform (HFD). The training data was used to retrain VGG16, VGG19, EfficientNetV2B3 and ViT-Vision Transformers based ViT-Base-Patch16-224-in21k models using transfer-based deep learning techniques. Models were tested with test data and the person's voice was classified according to whether it belonged to MS and other neurological diseases or a healthy individual. With this study, deep learning techniques were used for voice analysis in the diagnosis of neurological diseases. An overall accuracy value of over 96% in all four models used shows that this method can be used for neurological disease diagnosis. This provides opportunities for better understanding and treatment of neurological diseases by providing a basis for the discovery of new findings and methods. This thesis study offers a new approach to neurological disease diagnosis and aims to support efforts in the diagnosis and treatment of neurological diseases.