Automatic Assessment of Functional Movement Screening Exercises with Deep Learning Architectures.

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Machine Learning in Healthcare

Background:

• Effective physiotherapy requires accurate, unsupervised home exercise performance.
• Automated exercise evaluation can enhance therapy effectiveness and reduce injury risk.
• Existing methods underutilize deep learning's potential for motion analysis.

Purpose of the Study:

• To develop and evaluate a deep learning system for automatic physiotherapy exercise assessment.
• To investigate the system's performance using inertial measurement units (IMUs).
• To explore the potential of transfer learning for improving generalization.

Main Methods:

• Recorded four Functional Movement Screening (FMS) exercises using 17 IMUs.
• Trained a convolutional, long-short-term memory, and dense neural network architecture.
• Conducted extensive hyperparameter optimization and compared CNN structures for IMU data.

Main Results:

• The deep learning model accurately classified unknown exercise repetitions from known subjects.
• Consistent performance was not achieved on data from previously unknown subjects.
• Model performance varied across different FMS exercises.

Conclusions:

• Deep learning effectively performs complex motion analysis using IMU data.
• Performance on new subjects mirrors classical machine learning limitations.
• Transfer learning offers a promising solution for retraining and adapting the model.