Automatically Evaluating Balance: A Machine Learning Approach.

Area of Science:

• Biomechanics
• Machine Learning
• Rehabilitation Science

Background:

• In-home balance training is less effective than in-clinic due to limited physical therapist (PT) feedback.
• Objective, automated balance assessment tools are needed for effective remote rehabilitation.

Purpose of the Study:

• To assess the feasibility of using trunk sway data and machine learning (ML) for automated balance evaluation during standing exercises.
• To develop an ML model that can accurately map trunk sway features to physical therapist (PT) ratings of balance performance.

Main Methods:

• Sixteen participants performed standing balance exercises, with trunk sway data recorded.
• A physical therapist (PT) rated balance performance (1-5 scale); a 61-dimensional feature vector was extracted from sway data.
• A multi-class support vector machine (SVM) was trained to classify balance performance based on trunk sway features.

Main Results:

• The SVM model achieved 82% classification accuracy in a leave-one-participant-out validation scheme.
• ML-based assessments were significantly closer to PT ratings than participant self-assessments.
• Trunk sway analysis with ML shows promise for objective balance assessment.

Conclusions:

• Machine learning techniques can provide accurate, automated balance assessments during standing exercises without direct PT supervision.
• Automated balance assessment can potentially reduce PT consultation time and improve adherence to home-based rehabilitation programs.