Months-long High-performance Fixed LSTM Decoder for Cursor Control in Human Intracortical Brain-computer Interfaces

Abstract

Intracortical brain-computer interfaces (iBCIs) enable high performance cursor control for people with tetraplegia by inferring motor intentions from neural recordings. However, current methods rely on frequent decoder recalibrations to reduce performance fluctuations attributable to instability in neural recordings. Towards clinical translation, iBCIs must sustain high performance over long periods of time with minimal interruptions to the user. Recent non-human primate (NHP) studies indicate that recurrent neural network (RNN) decoders are more robust to neural variability. Here, we demonstrate that an RNN variant, a long short-term memory (LSTM) neural decoder, provides online long-term, stable two-dimensional cursor control for a participant with tetraplegia enrolled in the BrainGate2 clinical trial. An LSTM decoder was trained with multiple days of the participant's historical intracortical motor cortex recordings spanning seventy days. The LSTM decoder was then fixed and evaluated online as the participant used the iBCI to control a computer cursor during a center out and back task for 15 sessions across four months. The LSTM demonstrated high performance for the first three months without recalibration or adaptive parameter updates with an average performance of 93.8% of targets acquired. This longitudinal study suggests that a nonlinear RNN-based decoder can provide stable, intuitive control of 2-D kinematics by humans with tetraplegia.