Parkinson's Disease Classification using Pitch Synchronous Speech Segments and Fine Gaussian Kernels based SVM

Area of Science:

• Biomedical Engineering
• Speech Processing
• Neurology

Background:

• Conventional methods for Parkinson's disease (PD) speech analysis use fixed-time window segmentation.
• These methods extract features for classifying speech from PD patients versus healthy controls (HC).
• Limitations exist in accurately capturing speech dynamics with fixed windows.

Purpose of the Study:

• To evaluate pitch synchronous segmentation for improved PD speech classification.
• To compare the efficacy of pitch synchronous versus fixed window segmentation.
• To assess the impact of different vowel sounds on classification accuracy.

Main Methods:

• Mel-frequency cepstral coefficients (MFCCs) were extracted from pitch synchronous and fixed (25ms) window speech segments.
• Classification was performed using fine Gaussian support vector machines (SVM).
• Principal Component Analysis (PCA) and 10-fold cross-validation were employed for dimensionality reduction and performance evaluation.

Main Results:

• Pitch synchronous segmentation demonstrated superior classification performance compared to fixed window segmentation.
• Analysis of features grouped by vowel content provided insights into sound-specific effects.
• Clustering experiments successfully assigned class labels (PD/HC) based on participant distribution.

Conclusions:

• Pitch synchronous segmentation is more effective for classifying connected speech in Parkinson's disease.
• The proposed automatic speech analysis framework highlights the clinical relevance of pitch synchronous methods.
• This approach offers a more efficient alternative to traditional fixed-window techniques for PD assessment.