Multi-Channel Spectro-Temporal Representations for Speech-Based Parkinson's Disease Detection.

Area of Science:

• Computational Neuroscience
• Speech Processing
• Machine Learning for Healthcare

Background:

• Early detection of Parkinson's Disease (PD) is crucial for effective management and treatment.
• Speech analysis presents a promising avenue for non-invasive and scalable PD screening.
• Sentence-level speech analysis for PD detection remains an underexplored but clinically relevant area.

Purpose of the Study:

• To propose and evaluate a multi-channel spectro-temporal deep-learning approach for PD detection.
• To investigate the efficacy of fusing complementary time-frequency speech representations.
• To compare the performance of different deep learning architectures on this task.

Main Methods:

• Extracted and fused three time-frequency representations: mel spectrogram, Constant-Q Transform (CQT), and gammatone spectrogram.
• Created a three-channel input analogous to an RGB image for deep learning models.
• Evaluated Convolutional Neural Networks (CNNs) and Vision Transformers on the PC-GITA dataset using 10-fold subject-independent cross-validation.

Main Results:

• Multi-channel fusion consistently improved performance across all evaluated architectures compared to single representations.
• EfficientNet-B2 achieved the highest accuracy (84.39%) and F1-score (84.35%), outperforming recent methods.
• Performance varied with sentence type; emotionally salient and prosodically emphasized utterances showed higher discriminability.

Conclusions:

• Multi-channel spectro-temporal fusion enhances sensitivity to subtle speech impairments in Parkinson's Disease.
• The proposed deep learning approach offers a robust framework for speech-based PD screening.
• Further validation is required for potential clinical application of this speech analysis technique.