Sensitive Quantification of Cerebellar Speech Abnormalities Using Deep Learning Models

Area of Science:

• Neurology
• Speech Science
• Artificial Intelligence

Background:

• Cerebellar ataxias manifest early with detectable speech changes.
• Accurate disease assessment is crucial for clinical trials and patient care.
• Current speech assessments may lack sensitivity for early or subclinical disease stages.

Purpose of the Study:

• To develop and validate AI models for identifying and quantifying ataxic speech.
• To assess the models' ability to distinguish ataxia patients from healthy individuals.
• To evaluate the models' capacity for estimating disease severity and progression.

Main Methods:

• Utilized convolutional neural networks (CNNs) on log-mel spectrograms of speech.
• Employed time and frequency partial derivatives to capture motor speech phenotypes.
• Trained classification models for patient vs. control discrimination.
• Trained regression models for disease severity estimation.

Main Results:

• CNNs accurately classified healthy controls and ataxia patients, including those with no clinically detected speech deficits.
• Regression models provided precise disease severity estimates.
• Models detected subclinical speech alterations and tracked disease progression over time.

Conclusions:

• AI models analyzing speech signal derivatives can detect subclinical ataxic speech changes.
• These models offer sensitive, low-burden tools for early detection and monitoring in cerebellar ataxias.
• The approach holds potential for enhancing clinical trials and neurological care.