Exploring Speech Biosignatures for Traumatic Brain Injury and Neurodegeneration: Pilot Machine Learning Study.

Area of Science:

• Neuroscience
• Computational Linguistics
• Biomedical Engineering

Background:

• Speech features are increasingly recognized as potential biomarkers for neurodegenerative and mental health conditions.
• Early detection and differentiation between disorders through speech analysis are critical for effective diagnosis and management.

Purpose of the Study:

• To explore speech biosignatures in mild traumatic brain injuries (concussions) and Parkinson disease (PD).
• To evaluate the effectiveness of speech analysis in differentiating between these neurodegenerative conditions and healthy controls.

Main Methods:

• Utilized speech samples from participants with concussions, PD, and age-matched healthy controls for the PaTaKa and Sustained Vowel (/ah/) tests.
• Employed machine learning models (SVM, decision tree, random forest, XGBoost) with 37 temporal and spectral speech features.
• Applied data augmentation and 5-fold cross-validation to assess classification performance.

Main Results:

• The PaTaKa test achieved high F1-scores (>0.9) for classifying concussed vs. healthy and concussed vs. neurodegenerative conditions.
• Initial neurodegenerative vs. healthy classification showed poor performance (<0.2 F1-score), improved to 60-70% accuracy after data augmentation.
• The Sustained Vowel test demonstrated high F1-scores (>0.85) for concussed vs. neurodegenerative but lower scores for other comparisons.

Conclusions:

• Speech features hold potential as biomarkers for neurodegenerative conditions.
• The PaTaKa test shows strong discriminative ability, particularly for concussion-related and differential diagnoses.
• Further research is needed to refine speech-based tools for accurate neurodegenerative disease identification and differential diagnosis.