Estimating upper-extremity function from kinematics in stroke patients following goal-oriented computer-based training

Introduction After a stroke, a wide range of deficits can occur with varying onset latencies. As a result, assessing impairment and recovery are enormous challenges in neurorehabilitation. Although several clinical scales are generally accepted, they are time-consuming, show high inter-rater variability, have low ecological validity, and are vulnerable to biases introduced by compensatory movements and action modifications. Alternative methods need to be developed for efficient and objective assessment. In this study, we explore the potential of computer-based body tracking systems and classification tools to estimate the motor impairment of the more affected arm in stroke patients. Methods We present a method for estimating clinical scores from movement parameters that are extracted from kinematic data recorded during unsupervised computer-based rehabilitation sessions. We identify a number of kinematic descriptors that characterise the patients’ hemiparesis (e.g., movement smoothness, work area), we implement a double-noise model and perform a multivariate regression using clinical data from 98 stroke patients who completed a total of 191 sessions with RGS. Results Our results reveal a new digital biomarker of arm function, the Total Goal-Directed Movement (TGDM), which relates to the patients work area during the execution of goal-oriented reaching movements. The model’s performance to estimate FM-UE scores reaches an accuracy of R^2 : 0.38 with an error ( σ : 12.8). Next, we evaluate its reliability ( r=0.89 for test-retest), longitudinal external validity ( 95% true positive rate), sensitivity, and generalisation to other tasks that involve planar reaching movements ( R^2 : 0.39). The model achieves comparable accuracy also for the Chedoke Arm and Hand Activity Inventory ( R^2 : 0.40) and Barthel Index ( R^2 : 0.35). Conclusions Our results highlight the clinical value of kinematic data collected during unsupervised goal-oriented motor training with the RGS combined with data science techniques, and provide new insight into factors underlying recovery and its biomarkers.