Non-Invasive Voice-Based Early Detection of Parkinson’s Disease via Spectral Feature Analysis and Machine Learning Techniques

Abstract

Background: Parkinson’s disease (PD) is a chronic, slowly progressive, and irreversible neuropathological disorder characterized by the progressive degeneration of specific neurons responsible for producing neurotransmitters essential for motor control. Although PD primarily affects motor function, various non-motor symptoms commonly emerge during the prodromal phase. These include autonomic dysfunction, cognitive and neurobehavioral disorders, and sensory and sleep abnormalities. Notably, speech and voice alterations, particularly hypokinetic dysarthria, are frequent manifestations. This research presents a methodology to distinguish between individuals with PD and healthy controls using voice signals through speech recognition and machine learning (ML) techniques. A dataset comprising 81 voice samples (41 healthy controls and 40 PD patients) was utilized to extract two types of cepstral features: Mel-frequency cepstral coefficients (MFCCs) and subband-based cepstral coefficients (SBCs). These extracted features were used to train and evaluate three ML algorithms: Random Forest (RF), K-Nearest Neighbors (KNN), and Support Vector Machines (SVM). Results: Among the algorithms evaluated, the SVM-SBC model exhibited the highest performance, achieving an accuracy of 79%, a sensitivity of 75.5%, and an Area Under the ROC Curve (AUC-ROC) of 84%. Conclusions: This study highlights the potential of integrating cepstral features with machine learning algorithms to develop reliable, non-invasive tools for the early detection of PD.