PhD Thesis Defense: Juhyeon Lee, Advancing Objective Motor Severity Assessment in Cerebellar Ataxias Using Sensors and Machine Learning
Content
Speaker:
Abstract:
Cerebellar ataxias are a heterogeneous group of neurological disorders characterized by impaired coordination and balance due to dysfunction of the cerebellum and its associated pathways. As most ataxias are progressive, frequent and objective assessment of motor severity is essential for monitoring disease progression and evaluating therapeutic interventions. The lack of disease-modifying treatments further underscores the need for sensitive and frequent assessment to support clinical trials. However, current assessments rely on clinician observation–based rating scales, which limit the frequency, sensitivity, and objectivity of disease severity evaluation.
To address these limitations, various sensors, including wearable inertial sensors and cameras, have been suggested as promising tools for continuous and objective assessment of motor impairments. This dissertation develops analytic pipelines and machine learning approaches to estimate ataxia severity from inertial sensor and video data collected during upper- and lower-limb motor tasks.
The first part presents an analytic pipeline to estimate motor severity using ankle-worn inertial sensors during a gait task. By analyzing sub-second movement profiles and extracting submovement features, supervised learning models are trained to estimate clinician-rated severity. The pipeline is further extended to estimate ataxia severity in children with Ataxia-Telangiectasia while minimizing the effects of age-dependent immature motor characteristics.
The second part introduces a supervised contrastive learning framework for estimating ataxia severity using wrist-worn inertial sensors during upper-limb tasks. By leveraging pairwise relationships between samples, this approach captures relative differences in motor impairment and reduces reliance on subjective clinical scores, improving robustness to label noise.
The final part extends this framework to simultaneously acquired inertial sensor and video data during upper-limb motor tasks. Digital measures of ataxia severity are derived independently from each modality and compared to evaluate cross-modality agreement. Both modalities show strong associations with clinician-rated severity, and their estimates are highly concordant, indicating they capture a consistent representation of motor impairment.
Overall, this dissertation establishes a progression from feature-based modeling to contrastive learning and cross-modal evaluation, providing a framework for developing and validating digital measures of ataxia. These results support the use of wearable sensors and video as objective, scalable tools for monitoring disease severity in clinical and real-world settings.
Advisor:
Ivan Lee