Diagnosis Cerebellar Ataxia using Deep Learning with Time Series Transformed Image

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Cerebellar Ataxia (CA) impairs movement coordination, affecting balance, gait, and limb control.
• Early diagnosis of CA is crucial for timely rehabilitation and therapeutic interventions.

Purpose of the Study:

• To develop a simple, quantitative assessment framework for Cerebellar Ataxia using a motion sensor.
• To investigate the utility of a chest-worn kinematic sensor for evaluating balance capabilities in CA patients.

Main Methods:

• A pendant-worn kinematic sensor was utilized to collect movement data during the Romberg's standing test.
• Transformed-image techniques (recurrence plot, melspectrogram, Poincaré plot) were applied to sensor data.
• Deep learning models, specifically MobileNetV2, were employed for CA diagnosis using transformed images.

Main Results:

• The melspectrogram transform technique, combined with MobileNetV2, achieved the highest diagnostic performance.
• An average validation accuracy of 89.99% was obtained for Cerebellar Ataxia diagnosis.
• The motion sensor system demonstrated potential for early symptom detection and quantitative assessment.

Conclusions:

• A wearable motion sensor system offers a feasible approach for quantitative assessment and early diagnosis of Cerebellar Ataxia.
• Melspectrogram transformation coupled with deep learning shows promise for enhancing CA diagnostic accuracy.
• This framework can support personalized rehabilitation and therapeutic strategies for individuals with Cerebellar Ataxia.