Inferring Grasp Intentions from Arm Trajectories via Deep Learning to Enable Functional Movement in Quadriplegia Revised Title Determining Grasp Selection from Arm Trajectories via Deep Learning to Enable Functional Hand Movement in Tetraplegia

Background: Cervical spinal cord injury severely affects grasping ability of its survivors. Fortunately, many individuals with tetraplegia retain residual arm movements that allow them to reach for objects. We propose a wearable technology that utilizes arm movement trajectory information and deep learning methods to determine grasp selection. Furthermore, we combined this approach with neuromuscular stimulation to determine if self-driven functional hand movement could be enabled in spinal cord injury participants.Methods: Two cervical SCI participants performed arbitrary and natural reaching movements toward target objects in three-dimensional space, which were recorded using an inertial sensor worn on their wrist. Time series classifiers were trained to recognize the trajectories using either a Dynamic Time Warping (DTW) algorithm or a Long Short-Term Memory (LSTM) recurrent neural network. As an initial proof-of-concept, we demonstrate real-time classification of the arbitrary movements using DTW only (due to its implementation simplicity), which when used in combination with a high density neuromuscular stimulation sleeve with textile electrodes, enabled participants to perform functional grasping.Results: Participants were able to consistently perform arbitrary two-dimensional and three-dimensional arm movements which could be classified with high accuracy. Furthermore, it was found that natural reaching trajectories for two different target objects (requiring two different grasp types) were distinct and also discriminable with high accuracy. In offline comparisons, LSTM (mean accuracies 99%) performed significantly better than DTW (mean accuracies 86% and 83%) for both arbitrary and natural reaching movements, respectively. Type I and II errors occurred more frequently for DTW (up to 60% and 15%, respectively), whereas it stayed under 5% for LSTM. Also, DTW achieved online accuracy of 79%. Conclusions: We demonstrate the feasibility of utilizing arm trajectory information to determine grasp selection using a wearable inertial sensor along with DTW and deep learning methods. Importantly, this technology can be successfully used to control neuromuscular stimulation and restore functional independence to individuals living with paralysis.