A simulation study on the effects of neuronal ensemble properties on decoding algorithms for intracortical brain–machine interfaces

BACKGROUND: Intracortical brain–machine interfaces (BMIs) harness movement information by sensing neuronal activities using chronic microelectrode implants to restore lost functions to patients with paralysis. However, neuronal signals often vary over time, even within a day, forcing one to rebuild a BMI every time they operate it. The term “rebuild” means overall procedures for operating a BMI, such as decoder selection, decoder training, and decoder testing. It gives rise to a practical issue of what decoder should be built for a given neuronal ensemble. This study aims to address it by exploring how decoders’ performance varies with the neuronal properties. To extensively explore a range of neuronal properties, we conduct a simulation study. METHODS: Focusing on movement direction, we examine several basic neuronal properties, including the signal-to-noise ratio of neurons, the proportion of well-tuned neurons, the uniformity of their preferred directions (PDs), and the non-stationarity of PDs. We investigate the performance of three popular BMI decoders: Kalman filter, optimal linear estimator, and population vector algorithm. RESULTS: Our simulation results showed that decoding performance of all the decoders was affected more by the proportion of well-tuned neurons that their uniformity. CONCLUSIONS: Our study suggests a simulated scenario of how to choose a decoder for intracortical BMIs in various neuronal conditions.