Home-based rehabilitation is a promising approach for patients with mobility impairments or living in remote areas. However, this approach is still limited due to high cost, lack of personnel and scarcity of new technologies that could remotely monitor while efficiently interacting with patients. Low-cost wearable sensors, e.g. accelerometers, allow to monitor motor performance but their clinical adoption is still limited because of poor capacity of quantifying movement, inability to extract the context where movement occurs, difficulty in providing meaningful patient-tailored real-time feedback to follow progresses . These limitations could be addressed with a carefully designed artificial intelligence (AI)-based solution : we have developed a novel platform composed of wearable inertial sensors and AI algorithms, able to interact with the patients and monitor their performance in the unsupervised home environment. The aim of present work is to validate the whole platform on a large dataset of healthy subjects. Particularly, we analyze the ability of the AI to model and recognize a wide spectrum of rehabilitation exercises with minimum a priori knowledge provided by the user.
The platform consists of: 3 inertial sensors, 2 of which enclosed in bracelets for upper and lower limb, and one in a necklace for postural monitoring; a mobile device with simple and straightforward user interface; an embedded system, for safe data storage and computationally demanding operations. The user interacts with the platform, performing different sets of exercises while the AI processes low-level data coming from the sensors. The system collects relevant information about the motor performance and provides the user with motivating and engaging real time feedbacks. A dataset of 126 healthy subjects has been used to test the platform and validate the AI on 40 heterogeneous exercises commonly used for upper and lower limb rehabilitation.
The platform proved to be robust and reliable in collecting, synchronizing, storing and ensuring the usability of data (1 failure on 126 recordings). The AI showed high selectivity in recognizing the different exercises (large clusters color and number - coded in Figure) performed randomly by all the subjects, as well as in identifying the persons and the single exercise repetitions (shown in the inlet).
These promising results highlight that the platform is able to precisely model and monitor motor performance. To confirm its usefulness in clinics, a further validation is currently being performed on motor impaired patients. This will assess the ability to monitor and personalize the rehabilitation routine for each patient covering the specific needs of home-based rehabilitation. Finally, the platform results also in an affordable system, providing highly intelligent care in a cost-effective way.
MAGIC PCP Project N. 687228
 Hayward KS et al. 2016
 Lara OD and Labrador MA 2013