Author(s): MATSUDA, R., HAYAKAWA, M., KIMURA, A., Institution: HOSEI UNVERCITY, Country: JAPAN, Abstract-ID: 2436

INTRODUCTION:
Skilled players can throw balls or darts toward a target more accurately. How this accuracy is acquired remains an important question in motor learning. Many studies reported that release parameters converge to the solution manifold through practice, even when individual parameter values vary across trials (Sternad et al., 2011). In other words, they acquire coordination among release parameters. Release parameters are closely related to body movements because the projectile remains in contact with the hand until immediately before release. If release parameters can be calculated from body movements, it becomes possible to evaluate inter-joint coordination to reduce variability in arrival locations and its acquisition. Therefore, this study aims to construct a model that derives release parameters from body movements. However, the relationship between body movements and release parameters is unlikely to be uniform in high degrees of freedom (DOF). Improving model accuracy becomes difficult when such heterogeneous relationships coexist. Constructing separate models for each distinct relationship may improve predictive accuracy. Therefore, we applied a clustering analysis prior to constructing the statistical model and examined how the cluster changed with practice.
METHODS:
41 novice males threw darts aiming at the dartboard center. They threw 60 darts per set, and 5 sets per day for 10 days. A motion capture system recorded the kinematics of the body and the dart. The positions, velocities, and angular velocities of the upper extremity joints, and the postures of the upper extremity segments at the release timing were used as independent variables to estimate the release parameters. The Fisher-EM clustering algorithm was used to classify observations. After clustering, ridge regression, Gaussian process regression, and multilayer perceptron (MLP) models were constructed for each cluster and compared with respect to predictive accuracy.
RESULTS:
30 clusters were detected in the clustering analysis, and many subjects remained in the same cluster throughout the learning period. MLP had the highest prediction accuracy for all clusters. The root mean square error in the arrival locations derived from the estimated release parameters obtained from this model and the measurement observations was less than 10 cm.
CONCLUSION:
The clustering analysis results differed from the previous study, which found that arm movement trajectories formed multiple clusters during the early stages of learning but converged into a single cluster in the later stages of learning in a reaching task (Annada et al., 2022). This study implies that learning may not converge toward a single universal solution in a throwing task with high DOF. Additionally, this study showed a model with high predictive accuracy. This model enables evaluation of integrated coordination between body movements and release parameters, providing new insight into how motor redundancy is managed during skill acquisition.