Domain Adaptation for IMU Data to Enhance Objective Assessment of Friedreich Ataxia

Friedreich Ataxia (FRDA) is an uncommon progressive neurodegenerative disorder of the nervous system that primarily affects motor coordination. Measuring disease severity is crucial, as it allows clinicians to monitor disease progression and make appropriate treatment decisions. By using inertial measurement units (IMUs) to capture the motion of individuals with FRDA, it has been possible to develop algorithms that objectively assess motor impairment. These studies have primarily employed conventional machine learning (ML) methods but have been limited by small datasets, which is an inherent challenge in researching rare diseases. To address this, we propose a robust deep learning framework that integrates heterogeneous data sources to enhance model performance. Our approach leverages a convolutional neural network (CNN) architecture to automatically learn high-level representations from raw IMU signals, minimizing reliance on manual feature engineering. Central to our method is a two-stage training strategy that incorporates domain adversarial learning, enabling knowledge transfer between two IMU-based assessment tools: the Ataxia Instrumented Measures cup (AIM-C) and spoon (AIM-S). This strategy enhances learning from each device by exploiting shared underlying representations. The effectiveness of the framework was validated by comparing the coefficient of determination between clinical scores (the modified Friedreich's Ataxia Rating Scale) and model outputs from both standard CNNs and the proposed two-stage training approach. For the standard CNN models, the coefficients of determination were 0.53 for AIM-C and 0.46 for AIM-S. With the proposed two-stage training, these increased to 0.62 and 0.49, respectively, demonstrating clear performance gains through cross-device knowledge transfer.