State-dependent and context-aware intracortical neural decoding of motor intentions

The human brain functions as a central hub for orchestrating complex motor behaviors. Neurological disorders, injuries, and aging can impair motor functions, severely limiting autonomy and quality of life. Intracortical brain-computer interfaces (BCIs), which decode neural signals from microelectrode arrays implanted in the cortex, offer a promising avenue to restore lost functions through real-time control of external devices. Yet, a major challenge remains in translating laboratory successes into robust performance in real-word scenarios. Traditional BCI research often relies on highly controlled and stereotyped contexts, overlooking the dynamic, context-dependent nature of the real world. To address this, data from a neurobehavioral platform enabling wireless recordings from premotor neurons of macaques in both head-restrained (RCs) and freely moving contexts (FMCs) were analyzed in a first work exploring context-dependent neural activity. Neurons frequently exhibited different tuning across contexts, with rich mixed-selectivity encoding and distal-proximal motor synergies observed exclusively in the FMCs. Decoding across contexts further confirmed the superior generalizability of neural representations from FMCs, advocating for a neuroethological shift. Motivated by the presence of anticipatory and structured dynamics in premotor activity, this thesis also presents a decoding algorithm that expresses neural activity relative to stereotypical spike patterns from each neuron. Applied to macaque data in the RC, the method decoded hand and mouth movements preserving good accuracy even with large jittering and detected intentions over 500 ms before movement onset, underscoring its utility for prosthetic control. Building on these insights into context-dependent neural encoding, a second work extended the investigation to humans. In a clinical study involving a participant with amyotrophic lateral sclerosis, speech-related neural activity was examined during both listening and attempted speech to explore the potential of auditory signals to enhance communication BCIs. Passive listening posed minimal privacy concerns, while active listening enabled high decoding accuracy. Although cross-domain generalization between listening and speaking was limited, combining both modalities during training enhanced the decoding of attempted speech. Together, these findings advance the understanding of context-dependent neural encoding and propose strategies for developing adaptive, generalizable BCIs that can operate reliably in the complexity of everyday life.