Automatically evaluating balance using machine learning and data from a single inertial measurement unit.

Area of Science:

• Biomechanics
• Machine Learning
• Rehabilitation Engineering

Background:

• Machine learning (ML) is increasingly used for balance assessment with inertial measurement units (IMUs).
• Current ML methods often rely on hand-engineered features, which are labor-intensive and may not capture all relevant information.
• There is a need for automated feature extraction from IMU data for more efficient and potentially more accurate balance assessment.

Purpose of the Study:

• To investigate the utility of ML for automatically extracting features from raw IMU data for balance assessment.
• To compare the performance of ML models trained on unprocessed IMU data against models using hand-engineered features and self-assessments.

Main Methods:

• Ten participants with balance concerns performed balance exercises wearing an IMU on their lower back.
• Physical therapists rated balance performance on a 5-point scale via video recordings.
• ML models were trained using different representations of unprocessed IMU kinematic data and compared to models with hand-engineered features.

Main Results:

• ML models utilizing unprocessed IMU data achieved higher accuracy in estimating balance performance compared to models with hand-engineered features (AUROC 0.806 vs. 0.768).
• Performance using unprocessed data also surpassed participant self-assessments (AUROC 0.665).

Conclusions:

• Raw IMU data fed into ML models can accurately estimate balance performance.
• This approach offers a generalizable method for balance assessment, reducing the need for manual feature engineering while maintaining high performance.