ParkMAE: a cross-linguistic masked autoencoder framework for robust Parkinson's disease detection from speech

Area of Science:

• Neurology
• Speech Science
• Artificial Intelligence

Background:

• Parkinson's disease assessment relies on clinical evaluations, which can be subjective and time-consuming.
• Speech changes are early indicators of Parkinson's disease, offering a potential non-invasive biomarker.
• Existing speech analysis methods often lack robustness, multilingual capabilities, and comprehensive validation.

Purpose of the Study:

• To introduce ParkMAE, a multilingual speech foundation model for Parkinson's disease assessment.
• To establish a comprehensive benchmarking system for evaluating speech-based Parkinson's disease assessment tools.
• To explore the capabilities and limitations of current speech-based approaches for diagnosis, medication monitoring, disease staging, and cognitive assessment.

Main Methods:

• Curated large-scale multilingual speech datasets (750h pretraining, 100h evaluation) across four languages.
• Developed a self-supervised masked autoencoder (ParkMAE) pretrained on the multilingual corpus.
• Evaluated ParkMAE on cross-linguistic diagnosis, medication state monitoring, disease staging, and cognitive assessment (MoCA).

Main Results:

• ParkMAE achieved a 39% F1 score for cross-linguistic diagnosis, outperforming eGeMAPS and using 89% fewer parameters than Whisper.
• Demonstrated exceptional generalizability to unseen languages without finetuning.
• Showed predictive capability for cognitive assessment (MoCA, F1=0.56), but current methods struggle with medication monitoring and disease staging.

Conclusions:

• ParkMAE offers a robust, parameter-efficient, and multilingual foundation model for Parkinson's disease assessment.
• Speech-based assessment shows promise for diagnosis and cognitive monitoring but requires further development for medication state and disease staging.
• The study provides a reproducible framework for advancing speech-based Parkinson's disease research and clinical applications.