Empowering Upper-Limb Robot-Aided Rehabilitation with Artificial Intelligence



Robot-aided rehabilitation is a promising method for treating both musculoskeletal and neuromuscular disorders affecting the upper limb. These technologies can be coupled with conventional treatment to deliver highly intensive, specific sessions and improve the overall effectiveness of therapeutic interventions [1]. The integration of Artificial Intelligence (AI) algorithms in these rehabilitative robotic platforms may generate tailored and stimulating therapies, offering advantages for both patients and therapists [2]. This integration allows the systems to become capable of adjusting the behavior of robots within a therapy session based on quantitative measurements extracted from patients through multimodal monitoring interfaces that include physiological, kinetic, and kinematic recordings. Thus, the aim of this research project is exploring the use of different AI modules for robot-aided upper limb rehabilitation, with the goal of enhancing clinical outcomes and tailoring treatments to meet the users’ needs.  

Materials and Methods 

The proposed approach aimed to develop an autonomous strategy for dynamically adapting robot control parameters within the rehabilitative session replicating the decision-making process of the physiotherapist. To this end, it was necessary to set up an experimental protocol for acquiring labelled data, resulting in the development of a Physiotherapist-Informed Learning (PIL) paradigm.  

Moreover, the PIL was integrated in an experimental session to study the relationship between the physiotherapist decisions and the measurable parameters collected from the participants, via a multimodal monitoring interface, interacting with a 7 DoFs robot controlled with a tunable interaction control and equipped with an ergonomic flange. Ten orthopaedic patients, along with their therapists, were enrolled in the study involving a robot-aided session comprising nine repetitions of nine point-to-point movements within a 3D space. Physiotherapists real-time labelled the data with their decision to change the robot assistance level (i.e., raise/keep constant/lower the assistance level). Then, the obtained dataset was used to train and test different state-of-the-art Machine Learning algorithms. A k-fold cross-validation was applied on 70% of the dataset to compare algorithms accuracy, and the remaining 30% was used for hyperparameter optimization.  


This study revealed that the therapists’ choices are strictly related to the patients’ subjective perception of exertion and pain and their physical and cognitive workload. Conversely, the therapists’ choices exhibited a weak correlation with kinematic and kinetic data. Despite this, the best performance was obtained when the features fed into the AI modules belong to a multimodal monitoring. Specifically, the best optimized algorithm achieved 78.8±4.5% accuracy. Future studies are needed to: enlarge the dataset by involving more patients; clinically validate the AI-based adaptation approach. 


[1] C. Tamantini et al., “Tailoring Upper-Limb Robot-Aided orthopedic Rehabilitation on Patients’ Psychophysiological State”, IEEE Trans. Neural Syst. Rehabil. Eng., pp. 1–1, 2023, doi: 10.1109/TNSRE.2023.3298381. 

[2] D. Novak, M. Mihelj, J. Ziherl, A. Olenšek, e M. Munih, “Psychophysiological measurements in a biocooperative feedback loop for upper extremity rehabilitation”, IEEE Trans. Neural Syst. Rehabil. Eng. Publ. IEEE Eng. Med. Biol. Soc., vol. 19, fasc. 4, pp. 400–410, ago. 2011, doi: 10.1109/TNSRE.2011.2160357.