[Paper Review] Plug-and-Play Stability for Intracortical Brain-Computer Interfaces: A One-Year Demonstration of Seamless Brain-to-Text Communication

Chaofei Fan, Nick Hahn+7 more|PubMed|Nov 6, 2023

EEG and Brain-Computer InterfacesNeuroscience27 references19 citations

TL;DR

CORP enables continual online recalibration of an iBCI handwriting decoder using pseudo-labels corrected by a language model, achieving long-term stability over 403 days without supervised data.

ABSTRACT

Intracortical brain-computer interfaces (iBCIs) have shown promise for restoring rapid communication to people with neurological disorders such as amyotrophic lateral sclerosis (ALS). However, to maintain high performance over time, iBCIs typically need frequent recalibration to combat changes in the neural recordings that accrue over days. This requires iBCI users to stop using the iBCI and engage in supervised data collection, making the iBCI system hard to use. In this paper, we propose a method that enables self-recalibration of communication iBCIs without interrupting the user. Our method leverages large language models (LMs) to automatically correct errors in iBCI outputs. The self-recalibration process uses these corrected outputs ("pseudo-labels") to continually update the iBCI decoder online. Over a period of more than one year (403 days), we evaluated our Continual Online Recalibration with Pseudo-labels (CORP) framework with one clinical trial participant. CORP achieved a stable decoding accuracy of 93.84% in an online handwriting iBCI task, significantly outperforming other baseline methods. Notably, this is the longest-running iBCI stability demonstration involving a human participant. Our results provide the first evidence for long-term stabilization of a plug-and-play, high-performance communication iBCI, addressing a major barrier for the clinical translation of iBCIs.

Motivation & Objective

• Address nonstationarity in intracortical BCI decoding that necessitates frequent supervised recalibration.
• Develop a self-recalibration framework that operates online without interrupting user activity.
• Leverage language models to correct iBCI outputs and use pseudo-labels to retrain the decoder continually.
• Demonstrate long-term stability and quantify data requirements for seed model quality and recalibration.

Proposed method

• Introduce CORP: Continual Online Recalibration with Pseudo-labels for iBCIs.
• Use language models to automatically correct iBCI outputs and generate pseudo-labels.
• Apply online stochastic gradient descent to retrain the decoder using pseudo-labels with a replay buffer.
• Incorporate data augmentation (noise to features and mean offsets) to improve robustness.
• Employ a stopping criterion based on a loss threshold and learning rate to balance accuracy and recalibration time.
• Assess offline and online performance against baselines including no-recalibration and FA Stabilizer.

Experimental results

Research questions

• RQ1 Can CORP maintain high iBCI decoding accuracy over extended nonstationary operation without supervised retraining?
• RQ2 How does LM-assisted pseudo-labeling influence continual online recalibration performance?
• RQ3 What are the data requirements (seed data, recalibration samples) for effective long-term stability?
• RQ4 How does CORP compare to unsupervised distribution-alignment methods (e.g., FA Stabilizer) in handwriting iBCI?
• RQ5 What are the trade-offs between recalibration time and accuracy?

Key findings

• CORP maintained an average online handwriting decoding accuracy of 93.84% (±2.28) over 403 days.
• On day-0 CER was 3.22% for the seed model; with CORP, CER stayed low (6.35%±1.75% in online blocks) and stable across days.
• Offline WER for CORP was 3.68% (95% CI [3.21, 4.76]); online WER was 6.16% for CORP, versus 26.51% for no recalibration.
• CORP with ground-truth labels could reach 2.0% WER, illustrating the potential lower bound; CORP approaches this with pseudo-labels.
• LM-corrected pseudo-labels contributed most to recalibration accuracy; ablation without the LM raised CER to 19.22% (No Language Model).
• Recalibration requires only a small amount of data (≈10 sentences for substantial recovery); seed data from ~10 days yielded diminishing returns beyond that.