A Speech Signal Processing Approach for Early Detection of Parkinson's Disease

Sammanfattning

Parkinson’s disease (PD) is a progressive neurodegenerative disorder characterised by motor and non-motor symptoms, including significant speech impairments. Early and accurate diagnosis is crucial for effective management and treatment. However, traditional diagnostic methods are often limited because they rely on subjective assessments and only detect symptoms after they have become noticeable. This thesis explores the use of vowel articulation as a biomarker for detecting Parkinson’s disease, utilising acoustic analysis and machine learning techniques to create a predictive model. The study uses a dataset of speech recordings from individuals with and without PD, collected through Voice Diagnostics. Through a preprocessing and feature extraction process, relevant acoustic features are analysed. These features are then used to train a XGBoost machine learning model. The performance of the model is evaluated using metrics such as accuracy and the features were selected with the use of RFECV and Shapley scores. Results indicate model based on features from short form vowel extraction perform worse than previous work where longer vowel sounds were used. The model could not effectively distinguish between PD and non-PD speech patterns. The model achieves a moderately high classification accuracy on test data demonstrating the feasibility of this approach for PD diagnosis but lacks in reliability due to high variance in cross validation measurements. The findings suggest that, at this stage, integrating vowel sound analysis based on natural sentences into clinical practice does not improve early PD detection compared to previous work. Further work is required before this method can be used as a non-invasive and objective diagnostic tool. This research contributes to the growing body of literature on speech-based diagnostics and underscores the potential of machine learning in healthcare. Future work could expand on these findings by exploring larger datasets, additional acoustic features, and the application of this methodology to other neurological disorders.