Towards adaptive and finer rehabilitation assessment: A learning framework for kinematic evaluation of upper limb rehabilitation on an Armeo Spring exoskeleton

Abstract Providing specialized rehabilitation and tailoring the training process for patient’s needs and according to recovery potentials has gained importance. To satisfy this need, a dynamic assessment of the performance of the recovery process is required. Assessing rehabilitation for the upper limb is often carried out with clinical subjective scales that do not satisfy these requirements. The use of technologies introduced several sensors into the devices used for rehabilitation and permitted the rise of kinematic assessments. Kinematic measures provide an objective scale to follow up recovery during upper limb rehabilitation. The kinematics are still raw evaluations since they present insignificant effects if studied over short periods or on heterogeneous samples. We propose a framework for modeling the trajectories as a means of encoding the specificity of the movement at every stage. The new technique permits detecting significant differences as soon as three training sessions became available. We adopt an expectation–maximization algorithm and an optimization technique to encode the trajectories and the transition model from the acquired data. The framework enables us to encode in a Bayesian sense the observations from the patient and define six metrics to follow up on the progress of the movement quality. Statistical analysis of the results proved that these metrics are effective in tracking the evolution of the recovery. The results also established a strong discriminative property. The proposed framework promises a finer scale of evaluation and extends the knowledge about kinematic assessment. This study’s findings suggest that adopting these new metrics can help achieve more individualized patient care. It additionally promises to limit the amount of data needed to detect a significant change.