Brain-Computer Interfaces (BCIs) are an emerging approach to restore walking capabilities in people with paraplegia due to spinal cord injury (SCI) or other neurological conditions. BCIs record brain signals in real time to decode users’ intentions and use this information to operate external devices, such as orthoses, exoskeletons, or muscle stimulators. BCIs based on invasively recorded brain signals, such as electrocorticograms (ECoG), can achieve a better performance than their non invasive counterparts. This advantage comes from the fact that ECoG signals have higher spatiotemporal resolution than scalp recorded electroencephalograms (EEG). Moreover, subdurally implanted ECoG electrodes can elicit leg sensation by delivering cortical electrostimulation. Therefore, an ECoG-based BCI can restore both motor and sensory function to those paralyzed due to SCI. These so-called bi-directional BCIs (BD-BCIs) have the potential to be fully implantable, which greatly increases their potential for a widespread adoption. The main challenge in the development of fully-implantable BD-BCIs is complying with the FDA safety regulations for active implantable devices. These include protection from current leakage, thermal injury, and inflammatory responses, among others. In this work, we assessed the thermal safety of a fully-implantable ECoG-based BD-BCI system using the Finite Element Method (FEM). Specifically, starting from the FDA’s thermal safety constraints, we used bio-heat transfer models implemented in COMSOL to estimate the maximum power consumption (power budget) of the BD-BCI’s active components, namely, a chest wall unit (CWU) and a skull unit (SU). We also assessed the robustness of these computational models against the natural variation of physiological and environmental parameters, such as thermal, physiological, and metabolic properties of the tissues. Furthermore, we fabricated a CWU thermal prototype and performed benchtop experiments to validate our modeling approach. Finally, given that the decoder is one of the most “power-hungry” functional modules of the CWU, we developed a novel power-efficient decoding methodology capable of decoding individual steps with excellent accuracy and negligible lag. Specifically, we employed a combination of logistic regressions and the Fokker-Plank equation to design a recursive Bayesian filter that estimates the probability of leg swings from ECoG signals. We validated the performance of this decoder using ECoG signals recorded from motor cortical areas oftwo human subjects as they performed multiple walking tasks.
