Assessing Electromyographic and Kinematic Signals for Reach-and-Grasp Intention Decoding in Persons with Spinal Cord Injury

Human-machine interfaces (HMIs) based on muscular and kinematic information promise intuitive real-time control of assistive devices such as grasp neuroprosthesis for persons with cervical spinal cord injury (SCI). However, interpreting this data is challenging due to high dimensionality and nested discriminative information. Hence, feature engineering and ranking are imperative to minimize computational load while maintaining high performance. In this work, we recorded electromyography (EMG) and kinematic (acceleration, orientation, angular rate) information of inertial measurement units (IMUs) during reach-and-grasp movements (uni-/bimanual palmar/lateral grasps) in groups of non-disabled people (
$\mathrm{n}=12$
) and of people with incomplete cervical SCI (
$\mathrm{n}=3$
). We extracted 12 EMG and 11 IMU feature types of 8 EMG and 45 IMU channels. We applied the feature selection approaches chi-square, maximum-relevance-minimum-redundancy (MRMR), Random Forest (RF), and Boruta for dimensionality reduction of the feature set and evaluated resulting subsets. We could show for both groups that there was no significant decrease in classification accuracy (RF model) with chi-square and Boruta subsets compared to the baseline set of all features, despite their heavily reduced dimensionalities (<25% and <74%, respectively). Accuracies peaked at 98.6 ± STD 0.9% (control group, Boruta) and 97.7 ± STD 1.1% (participants with SCI, Boruta). We found the MRMR subsets to be performing significantly worse. We could further show high information interpretability of chi-square and RF scores that indicated the importance of sensors and extracted features for reach-and-grasp classification. We plan to investigate how the approach can be implemented in real-time reach-and-grasp HMIs for persons with SCI.