Developing a Real-Time Decoder for Bilateral Upper Extremity Motor Signal From Left Motor Strip ECoG After Cervical Spinal Cord Injury Using Multi-Class Machine-Learning Techniques

Introduction: Cervical spinal cord injury (SCI) incurs serious functional impairment in the upper and lower extremity with negligible prognosis for functional recovery. Previous investigations have explored volitional control restoration in one subject with complete cervical quadriplegia using a fully implanted electrocorticography (ECoG) device as a brain-computer interface (BCI). A decoder was trained using movement-intent of the dominant hand to control a mechanical hand orthosis. Expanding the decoder's ability to detect a greater degree of freedom of bimanual motor imageries using two 4-contact leads has yet to be explored whether this protocol can extrapolate to non-dominant extremity movement-intent or combined upper extremity movement-intent in a continuous multi-class decoding format.

Methods: Data were collected from 83 sessions from a 21-year-old male with complete cervical quadriplegia (C5 American Spinal Injury Association Impairment Scale A) who was previously implanted with two 4-contact ECoG leads over the dominant sensorimotor cortex in a prior clinical trial connected to a pulse generator capable of recording 2 time-series channels and 2 channels of average band power centered at 20Hz. Trials were labeled via motor-intent command and stored in BIDS format before data was windowed and converted to power spectral density. A gradient boosting classifier was trained on 80 sessions and validated using 5-fold cross-validation to predict left, right, both, or neither hand grasp motor intent. A real-time simulator was developed to predict outcomes from the remaining 3 validation sessions.

Results: An optimized gradient boosting algorithm predicted test classifiers with a mean AUC of 0.98 during 5-fold cross validation and a mean sensitivity, specificity, and F-1 Score above 0.80. Average single-session validation was 65.16%.

Conclusion : While more sophisticated optimizations will offer room for improvement, this study shows promise for bimanual motor-intent brain-computer interface using surgically-implanted ECoG devices to restore motor abilities after cervical SCI.