Automatic Evaluation of Motor Rehabilitation Exercises Based on Deep Mixture Density Neural Networks

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Rehabilitation Science

Background:

• Physical telerehabilitation systems offer potential for reduced treatment time and cost.
• Assessing physical rehabilitation automatically faces challenges due to movement complexities, nonlinearities, and stochastic uncertainties.
• Existing methods struggle to simultaneously address both the complexity of motion data and inherent stochastic uncertainties.

Purpose of the Study:

• To develop and evaluate a Deep Mixture Density Network (DMDN) for automatic physical telerehabilitation assessment.
• To improve the accuracy and generalization capabilities of automated movement analysis in telerehabilitation.
• To create a system that effectively handles the stochastic uncertainties and complexities of human motion data.

Main Methods:

• A Deep Mixture Density Network (DMDN) was proposed, combining deep neural networks with probabilistic models.
• A multi-branch convolutional layer extracted deep features, a Long Short-Term Memory (LSTM) network captured temporal dependencies, and a Gaussian Mixture Model (GMM) managed stochastic interactions.
• Input data consisted of joint position and orientation time series from Kinect v2, trained with clinical reference scores.

Main Results:

• The proposed DMDN, particularly with one-dimensional parallel window transitions, outperformed competing strategies in ablation studies.
• The DMDN demonstrated higher reliability, evidenced by a lower Root Mean Square Error (RMSE) standard deviation compared to a model without GMM.
• Performance was ranked competitively using Spearman correlation coefficient and RMSE against state-of-the-art algorithms.

Conclusions:

• The developed DMDN provides a robust framework for automatic physical telerehabilitation assessment, effectively integrating deep learning and probabilistic modeling.
• The DMDN architecture successfully addresses the complexities and stochastic nature of human movement data, leading to improved assessment validity.
• The findings suggest that DMDN offers a promising approach for enhancing the efficiency, accuracy, and reliability of telerehabilitation monitoring.