Sensitive quantification of cerebellar speech abnormalities using deep learning models.

Area of Science:

• Neurology
• Speech Science
• Artificial Intelligence

Background:

• Objective disease assessment is crucial for clinical trials and care.
• Cerebellar ataxias manifest early and obviously through speech changes.
• Developing accurate models to identify and quantify speech abnormalities in ataxia is essential.

Approach:

• Utilized deep learning models, specifically ResNet 18.
• Input features included time and frequency partial derivatives of log-mel spectrograms of speech.
• Trained models for classification (ataxia vs. healthy) and regression (disease severity estimation).

Key Points:

• The model accurately distinguished individuals with ataxia from healthy controls, including those with subclinical speech deficits.
• Regression models provided precise estimates of disease severity.
• The approach successfully measured subclinical signs of ataxia and captured disease progression over time.

Conclusions:

• Deep learning models analyzing speech signal derivatives can detect subclinical speech changes in ataxias.
• These models offer sensitive measurement of disease progression over time.
• Potential applications include early ataxia detection and low-burden assessment tools for clinical trials and neurological care.