Abstract
Balance assessment is crucial for health monitoring and rehabilitation evaluation of neurological diseases like Parkinson's disease (PD) and stroke. The Berg Balance Scale (BBS) is a widely used clinical tool for balance evaluation. However, its dependence on trained therapists for subjective, time-consuming assessments limits its scalability. Current researchers have proposed several automated assessment systems. However, they suffer from difficulty in use in clinical settings and the need for feature engineering. The rapid advancement of wearable inertial measurement units (IMUs) provides an objective tool for motion analysis that is suitable for use in clinical environments. Thus, to address the limitations of manual scoring and complexities of capturing gait features, we proposed an automated BBS assessment system using an attention-based deep learning algorithm with IMU data, integrating convolutional neural networks (CNNs) for spatial feature extraction, bidirectional long short-term memory (Bi-LSTM) networks for temporal modeling, and attention mechanisms to emphasize informative features. Validated with 20 healthy subjects (young and elderly) and 20 patients (PD and stroke), the system achieved a mean absolute error (MAE) of 1.1627 and root mean squared error (RMSE) of 1.5333. Requiring only 5 min of walking data, this approach provided an efficient, objective solution for balance assessment to assist healthcare physicians as well as patients in their own health monitoring. The key limitations included: a limited generalizability to severely impaired patients who were unable to walk independently, and the inability to predict the score of individual tasks.