Contrastive Learning Model for Wearable-Based Ataxia Assessment.

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Objective assessment of ataxia severity is crucial for disease tracking and treatment evaluation.
• Wearable sensors offer a promising avenue for frequent ataxia assessment.
• Current wearable methods often rely on features derived from subjective clinical scales, limiting accuracy and flexibility.

Purpose of the Study:

• To introduce a novel contrastive learning-based model for ataxia severity assessment using wearable inertial data.
• To develop a method that learns relevant features directly from motor severity differences, reducing bias from clinical scales.

Main Methods:

• Trained a contrastive learning model on inertial data from 87 ataxia patients and 44 healthy controls during a finger-to-nose task.
• Utilized a pairwise contrastive loss function to learn representations of relative ataxia severity.
• Evaluated learned features through downstream regression and classification tasks.

Main Results:

• Learned features showed strong associations with clinical scores (cross-sectional r=0.84, longitudinal r=0.68) and high reliability (ICC=0.96).
• The model achieved high accuracy in distinguishing between ataxia and healthy phenotypes (AUC=0.95).
• Outperformed existing methods relying on clinical score-derived features.

Conclusions:

• The contrastive model provides robust ataxia severity representations with less dependence on clinical scales.
• This approach enhances objectivity, scalability, and frequency of ataxia assessment.
• Potential to significantly improve patient monitoring and clinical trial efficiency in ataxia research.