Long-term unsupervised recalibration of cursor BCIs

Abstract

Intracortical brain-computer interfaces (iBCIs) require frequent recalibration to maintain robust performance due to changes in neural activity that accumulate over time. Compensating for this nonstationarity would enable consistently high performance without the need for supervised recalibration periods, where users cannot engage in free use of their device. Here we introduce a hidden Markov model (HMM) to infer what targets users are moving toward during iBCI use. We then retrain the system using these inferred targets, enabling unsupervised adaptation to changing neural activity. Our approach outperforms the state of the art in large-scale, closed-loop simulations over two months and in closed-loop with a human iBCI user over one month. Leveraging an offline dataset spanning five years of iBCI recordings, we further show how recently proposed data distribution-matching approaches to recalibration fail over long time scales; only target-inference methods appear capable of enabling long-term unsupervised recalibration. Our results demonstrate how task structure can be used to bootstrap a noisy decoder into a highly-performant one, thereby overcoming one of the major barriers to clinically translating BCIs.

Measuring instability in chronic human intracortical neural recordings towards stable, long-term brain-computer interfaces

“…To first establish a baseline for decoder performance, we deployed fixed decoders 27,51 for the purpose of identifying, over a comparatively long period, how neural instabilities may lead to deteriorating control. Data were collected from 15 consecutive research sessions spanning 142 days of T11 performing a center-out-and-back task using a fixed nonlinear (recurrent neural network) decoder, as previously described 51 (see Methods).…”

“…The type and magnitude of model drift results in various forms of performance degradation 19 , sometimes necessitating decoder recalibration to restore control. Existing solutions to reduce the need for recalibration tasks include adaptive decoders that require shorter recalibration sessions to maintain or restore stable performance 3,10,42 , self-supervised recalibration using retrospective labeling that avoids explicit recalibration sessions 26,27,43,44 , and robust decoders that experience less model drift by extracting stable, time-invariant features from high-dimensional recordings [20][21][22][45][46][47][48][49][50][51] or by adaptively adjusting decoder parameters 12,52 .…”

A real-time, high-performance brain-computer interface for finger decoding and quadcopter control

“…Although high performance was achieved using this decoding system, potential improvements to increase the likelihood of clinical adoption include reducing the calibration times and increasing the robustness to neural instabilities. Several approaches could be applied to this decoding system, including rapid decoder calibration 35 , training decoders using a long history of previously recorded data 36 , adaptive decoders using task knowledge 37,38 , and algorithms that perform dimensionality reduction to a stable manifold followed by realignment 39,40 .…”

Neural Data Transformer 2: Multi-context Pretraining for Neural Spiking Activity

“…Without mapping that range, BCIs based on neural data pretraining alone will need continuous supervision. This is still a practical path forward: beyond explicit calibration phases, strategies for supervising BCIs are diverse, and can derive from user feedback [59], neural error signals [60, 61], or task-based estimation of user intent [62, 57, 63]. The ambition of pretrained BCI models, with broad paired coverage of the neural-behavior domain will be challenging given the experimental costs of data collection; the field of robotics suggests that scaled offline or simulated learning are important strategies given this expense.…”

“…The ambition of pretrained BCI models, with broad paired coverage of the neural-behavior domain will be challenging given the experimental costs of data collection; the field of robotics suggests that scaled offline or simulated learning are important strategies given this expense. Since we lack convincing closed-loop neural data simulators (though see [64, 62]), understanding how to leverage behavior from existing heterogeneous datasets is an important next step.…”