An additional ally for physical rehabilitation: Artificial Intelligence
Every year millions of people undergo physical rehabilitation: on average, it takes weeks to restore physical fitness during which you have to perform rehabilitation exercises that last from half an hour to two hours.
90% of physical rehabilitation sessions are performed in a domestic context and this, according to some studies, involves deviations from the treatment prescribed by the doctor, with consequent dilation of recovery times and increase in health care costs.
Assessments of progress achieved in the physical rehabilitation phase are traditionally performed either manually or with the help of traditional computer systems that do not provide a type of feedback that can actually allow patients to understand where they are doing wrong and thus improve. University of Idaho researchers believe that artificial intelligence can help physicians follow their patients remotely, improving compliance with prescribed treatments.
Their work has been described in a document entitled ” Deep learning framework for the evaluation of physical rehabilitation exercises “. The framework is divided into three main points: metrics to quantify performance in exercises, conversion of performance within a scoring system and machine learning models that encode the relationship between movement data and quality scores.
Researchers collected skeletal data from motion exercises performed by 10 healthy patients through an optical tracking system. Then, they calculated two commonly used metrics for rehabilitation assessment: with models and without models.
In the first case, the repetitions are compared with exercise models, while in the second they are measurements coming directly from the trajectories carried out in the exercises. Subsequently, a further model was applied to reduce the number of variables that can affect the data and defined a scoring system that works in such a way that the value of the performance metrics is expressed through a quality score of the movement between 0 and 1.
Once the scores were obtained, the team formed three distinct neural networks: a convoluted neural network (CNN), a recurrent neural network (RNN), and a holistic network (HNN) composed of subnets. Each network has been ” trained ” five times and the team has verified the average deviation between the input quality scores and the expected quality scores from the network. After these studies, the CNN network proved to be the most accurate.
The team warns that the results are not necessarily generalizable, because the data set used was from healthy patients. However, they believe that this work can lay the foundation for further future applications. ”
To our knowledge, ours is the first work that implements neural networks for the evaluation of rehabilitation performance “, they say. ” Despite the essential role of rehabilitative assessment in improving rehabilitation outcomes and reducing healthcare costs, existing, approaches to computer-aided monitoring and patient performance assessment lack versatility, robustness, and practical relevance. “
