Semi-Supervised Classification for Intracortical Brain-Computer Interfaces Semi-Supervised Classification for Intracortical Brain-Computer Interfaces

Intracortical brain-computer interface (BCI) systems may one day allow paralyzed patients to interface with robotic arms or computer programs using their thoughts alone. However, a common and unaddressed issue with these systems is that due to small instabilities in the recorded signals, the decoding algorithms they rely upon must be retrained daily in a supervised manner. While this may be acceptable in a laboratory, it presents a burden to the patient. In this work, using data recorded from two subjects over a period of 41 (36) separate days, we first investigate the behavior of a standard decoder without daily retraining. Contrary to what might be expected, we find that mean daily accuracy does not significantly decline with time, though daily performance may be highly variable without retraining. Second, we investigate the behavior of the neural signals in both datasets across time and find that within day changes of the signal characteristics are largely dwarfed by between day changes. Finally, we present two algorithms which adapt to changes in neural signals in a semi-supervised manner and demonstrate stable decoding on both datasets without any daily supervised retraining.