Human pose estimation in physiotherapy fitness exercise correction using novel transfer learning approach.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Rehabilitation Science

Background:

• Physical therapy requires accurate monitoring of patient exercises for effective rehabilitation.
• Current methods may lack precision and accessibility, leading to suboptimal outcomes.
• Wearable sensor technology offers a promising avenue for objective exercise assessment.

Purpose of the Study:

• To develop and evaluate an efficient neural network approach for human pose estimation and correction during physical therapy.
• To utilize wearable sensor data for accurate classification of physical therapy exercises.
• To enhance rehabilitation monitoring through advanced machine learning techniques.

Main Methods:

• Leveraged a large dataset (276,625 records) from wearable inertial and magnetic sensors.
• Implemented a novel Random Forest Long Short-Term Memory (RFL) method integrating Long Short-Term Memory and Random Forest.
• Generated novel temporal and probabilistic features from smartphone sensor data for machine learning classification.

Main Results:

• The RFL approach achieved 99% accuracy in classifying physical therapy exercises.
• Rigorous experiments, including k-fold validation and hyperparameter optimization, confirmed the method's efficacy.
• The Random Forest component within RFL demonstrated superior performance.

Conclusions:

• The RFL method offers a novel feature generation approach that significantly improves exercise classification accuracy.
• This integration enhances rehabilitation monitoring and supports intelligent physiotherapy assistance systems.
• The technology presents a feasible alternative to frequent clinic visits, mitigating risks for patients with disabilities or major diseases.