Compensatory Movement Detection in Upper Limb Rehabilitation With Deep Learning Methods

Stroke risk increases with age. After the stroke, the patients are likely to develop some inappropriate compensatory movements. These movements disrupt normal activities and reduce the effectiveness of the rehabilitation. This paper applies deep learning algorithms like RNN, GRU, LSTM, and Transformer to detect the compensatory movements based on the Microsoft Kinect data from the Toronto rehab stroke pose dataset. The models are trained with focal loss to decrease the effect of the imbalanced data distribution. The trained models have achieved higher accuracy than works done in the literature on this dataset.