Early detection of parkinson's disease from voice data using deep learning approaches

Parkinson's disease is a neurodegenerative disorder affecting the motor system, caused by the degeneration of nerve cells in the brain that produce the neurotransmitter dopamine. Dopamine is responsible for the control, coordination, and fluidity of movements. Due to dopamine deficiency, patients develop motor symptoms such as tremors, bradykinesia (slowness of movement), muscle rigidity, and balance impairments. Although Parkinson's disease is commonly observed in older adults, it can also rarely occur at a younger age. The disease affects not only motor functions but also speech and voice. Changes in the brain can disrupt speech prosody. Common speech symptoms include soft or quiet speech, monotone intonation, and difficulties in articulation and slurring. These symptoms may hinder patients' effective communication with their surroundings. In this thesis study, novel artificial intelligence-based approaches employing machine learning and deep learning methods, including convolutional neural networks (CNNs), were developed for the detection of Parkinson's disease through voice-based biomarkers, providing new data to the literature. The study utilized voice signals obtained during the pronunciation of vowels (/a/, /o/, and /i/) and the Turkish word "gofret". The collected audio data were transformed into various visual representations spectrogram, chromogram, MFCC (Mel-Frequency Cepstral Coefficients), and tempogram and data augmentation techniques (time masking, stretching, shifting, and frequency masking) were applied to these images to perform binary classification tasks. Additionally, CNN models were developed for the classification of different stages of Parkinson's disease. According to the best results obtained in this study, the MIS-DL (Multi-vowel Input Spectrogram-based Deep Learning) model achieved an accuracy of 97.5% in distinguishing between diseased and healthy subjects. For stage classification, the PDVS-DL (Parkinson Disease Voice Stage-Deep Learning) model attained a performance of 93.1%.