Automated Assessment of Balance Rehabilitation Exercises With a Data-Driven Scoring Model: Algorithm Development and...

Area of Science:

• Rehabilitation Medicine
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Balance disorders are common, and traditional rehabilitation programs face barriers like resource limitations and therapist shortages.
• Home-based rehabilitation is challenging due to potential patient errors and lack of adherence.
• Holobalance offers a comprehensive home-based rehabilitation solution using persuasive coaching and augmented reality.

Purpose of the Study:

• To design, implement, and evaluate a data-driven scoring model for accurate assessment of balance rehabilitation exercises.
• To leverage machine learning for inferring exercise performance scores.

Main Methods:

• Utilized an extensive dataset of approximately 1300 rehabilitation sessions from the Holobalance pilot study.
• Trained machine learning models, including random forests and neural networks, on preprocessed, cleansed, and normalized data.
• Experts scored 1313 exercises based on a rubric to create a training and testing dataset.

Main Results:

• The data-driven scoring model demonstrated improved accuracy compared to a rule-based system, with k-statistic values ranging from 15.9% to 26.8%.
• Model performance approached interobserver variability thresholds, ensuring trustworthy scoring within the Holobalance system.
• Achieved high classification accuracy for sitting (0.86-0.90), standing (0.85-0.92), and walking (0.81-0.90) exercises.

Conclusions:

• Machine learning models effectively score balance rehabilitation exercises within the Holobalance system.
• The scoring module's accuracy, comparable to interobserver variability, allows reliable exercise assessment using sensor data.
• High classification accuracy supports the use of this model for home-based balance rehabilitation.