A deep learning system to monitor and assess rehabilitation exercises in home-based remote and unsupervised conditions

In physical rehabilitation, the progress in machine learning and the advent of affordable and low- cost motion capture solutions have been conducive to the development of approaches for automated monitoring of patient performance. A system for rehabilitation exercise repetition counting and validation based on a set of skeleton- based features that are obtained from a 2D human pose estimation network was proposed. A deep learning application was developed to provide real-time feedback with much more specific information regarding exactly how the motion deviates from the correct execution. To this end, a dataset of 6 resistance training exercises was acquired to train a deep neural network to predict the exercises' movements at a frame level. Hence, this underlying idea of inferring the moment of the exercise is two-fold: (i) to provide information about the exercise execution with a fine level of detail; (ii) the ability to detect invalid repetitions promptly. Finally, a repetition counting and a validation module receive the predicted moment and output the current number of valid repetitions. The proposed system showed good performance in motion analysis resulting in an unobtrusive monitoring and evaluation method that could be tested for rehabilitation scopes in-home based unsupervised settings. The development of systems that can reliably capture human movements, automatically analyze the recorded data and evaluate the quality of the movement performance will play an important role in supplementing traditional rehabilitation assessments performed by trained clinicians and assisting patients participating in home-based rehabilitation.