Deep Learning and Artificial Intelligence Applied to Model Speech and Language in Parkinson's Disease

Area of Science:

• Neuroscience
• Computational Linguistics
• Artificial Intelligence

Background:

• Parkinson's disease (PD) is a prevalent neurodegenerative disorder affecting motor and non-motor functions, including speech and language.
• While speech impairments in PD are studied, language-based biomarkers for cognitive assessment remain underexplored.
• Automatic detection and monitoring of PD using biomarkers is an active research area.

Purpose of the Study:

• To propose and evaluate automatic assessment methodologies for Parkinson's disease using speech and language biomarkers.
• To compare the effectiveness of different machine learning models, including CNNs and pre-trained networks, for PD classification.
• To investigate the impact of fusing speech and language modalities on classification accuracy.

Main Methods:

• Utilized one-dimensional and two-dimensional Convolutional Neural Networks (CNNs).
• Employed pre-trained models: Wav2Vec 2.0 for speech and BERT/BETO for language.
• Investigated independent modeling of speech and language, followed by early, joint, and late fusion strategies.

Main Results:

• Speech modality achieved up to 88% accuracy in classifying Parkinson's disease patients versus Healthy Controls (HC).
• Speech representations demonstrated superior discrimination compared to language representations and multi-modal approaches.
• Fusion strategies indicated potential information loss in the speech modality with changes in multi-modal representation time spans, impacting accuracy.

Conclusions:

• Speech biomarkers are more effective than language biomarkers for discriminating Parkinson's disease patients from healthy individuals.
• Further research is needed to explore optimal fusion methods and time spans for multi-modal analysis in PD detection.
• Automatic analysis of speech offers a promising avenue for the detection and monitoring of Parkinson's disease.