(L-461) Identifying independent and assisted step patterns from accelerometer data acquired during step sessions with task-specific spinal cord epidural stimulation of an individual with spinal cord injury.

Introduction:: Previous evidence has suggested that task-specific spinal cord epidural stimulation (scES) with locomotor training led to recovery of independent standing and walking ability for individuals with Spinal Cord Injury (SCI).¹ However, even after rehabilitation, the recovery is limited to irregular locomotor patterns with reduced speed.² Proper identification of such patterns and spatiotemporal parameters is important throughout rehabilitation. While predictive control of epidural stimulation is an unexplored area, predictive algorithms have been employed previously in FES-assisted locomotion of individuals with motor incomplete SCI using rule-based learning approaches.³ However, accuracy of rule-based learning was less than the Random Forest Classifier (RFC) model in identifying sitting, standing, walking, and running in able-bodied individuals.⁴ Hence, we aimed to use accelerometer data and the RFC model to identify independent and assisted stepping patterns in an individual with SCI.

Materials and Methods:: One participant with chronic motor complete SCI (AIS A-T3) with an implanted neurostimulator underwent 160 alternating sessions of stand and step training. Accelerometers were placed on the sacrum and bilaterally on the shank and the foot. A manual pulse was used to record changes in assistance needed during selected step training sessions with scES. Heel strike and toe-off events were identified for both legs using foot acceleration data. For each side, a 12-dimensional feature vector with foot acceleration and roll angle data was obtained for every gait cycle window. The RFC model was used to predict assistance or levels of step independence (swing, stance, or full gait cycle).


Results, Conclusions, and Discussions:: The prediction accuracy of the model was 82.61% (left leg) and 81.81% (right leg). A feature to identify assisted or independent swing patterns was the roll valley which was the duration between initial movement of the foot roll angle and the subsequent heel strike. It was 38.98 ± 8.87% of the gait cycle when fully assisted, and 25.71 ± 15.21% for independent swing. The two most prominent features for the left leg were variance of the roll (22.40%) and roll valley (16.60%) and for the right leg were mean of the vertical foot acceleration (18.93%) and stride time (17.67%). Hence, the RFC model with the roll angle pattern was able to identify levels of assistance. Using a machine learning approach to analyze daily training data provides greater insight on the progression of the level of independence achieved during stepping with scES.

Acknowledgements (Optional): :

References (Optional): : [1] Angeli, C. A., Boakye, M., Morton, R. A., Vogt, J., Benton, K., Chen, Y., ... & Harkema, S. J. (2018). Recovery of over-ground walking after chronic motor complete spinal cord injury. New England Journal of Medicine, 379(13), 1244-1250.

[2] Kececi, A., Yildirak, A., Ozyazici, K., Ayluctarhan, G., Agbulut, O., & Zincir, I. (2020). Implementation of machine learning algorithms for gait recognition. Engineering Science and Technology, an International Journal, 23(4), 931-937.

[3] Kostov, A., Andrews, B. J., Popovic, D. B., Stein, R. B., & Armstrong, W. W. (1995). Machine learning in control of functional electrical stimulation systems for locomotion. IEEE Transactions on Biomedical Engineering, 42(6), 541-551.

[4] van der Salm A., Nene, A. V., Maxwell, D. J., Veltink, P. H., Hermens, H. J., & IJzerman M. J. (2005). Gait impairments in a group of patients with incomplete spinal cord injury and their relevance regarding therapeutic approaches using functional electrical stimulation. Artif Organs. Jan;29(1):8-14. doi: 10.1111/j.1525-1594.2004.29004.x. PMID: 15644078.