A Personalized and Adaptable User Interface for a Speech and Cursor Brain-Computer Interface
Hamza Peracha, Carrina Iacobacci, Tyler Singer-Clark, Leigh R Hochberg, Sergey D. Stavisky, David M. Brandman, Nicholas S Card · 2026 · Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26) · doi:10.1145/3772318.3790750
Summary
This paper reports on the user interface design for an intracortical brain-computer interface (BCI) deployed for everyday at-home use by people with severe paralysis. The work centers on a 22-month longitudinal co-design study with one BrainGate2 trial participant, T15 - a 45-year-old man with tetraplegia and severe dysarthria due to ALS who has four 64-channel microelectrode arrays implanted in his speech motor cortex - and then demonstrates the same framework's adaptability with a second participant, T19, a 41-year-old man with tetraplegia and anarthria due to ALS, who has only forehead, eyebrow, and eye movements remaining and is ventilator-dependent. The team treated the BCI as a technology probe, deploying working prototypes in T15's home, holding twice-weekly check-ins, formal surveys every six months, and continuous text and shared-document collaboration. The system architecture decouples a finite-state-machine logic node from a Python pyglet graphics node, both running on the BRAND real-time platform, which allowed the same backend to drive separate frontends for the BCI workstation, an iPad, and a macOS desktop. Available input modalities include decoded speech, neural cursor control, decoded gestures, and eye tracking, with the user free to switch modalities depending on context. Speech decoding uses a transformer-based phoneme decoder feeding an n-gram model and a fully local OPT-6.7B large language model for rescoring; cursor and click decoders are lightweight linear models.
Key findings
T15 used the BCI for over 4,000 hours across 22 months, up to 19 hours per day, communicating at up to 60 words per minute and maintaining full-time employment, family conversations with his young child, email, video calls, web browsing, and report writing. Three iterations of the sentence-correction interface raised the percentage of sentences marked fully correct from 40-41% to 59%, while time spent correcting per trial increased from 19 to 34-62 seconds, indicating users were willing to trade speed for accuracy when the workflow supported word-level edits. T15 used eye tracking for 88.5% of interface time (intuitive, large-button magnetized layout) and neural cursor for 11.5% (necessary for personal-computer control where buttons are not BCI-optimized, and more reliable in poor lighting). Word-level correction handled 91.2% of corrected sentences; manual on-screen-keyboard typing was the most accurate correction (76% fully correct) but slowest. User-centered design surveys at 285, 469, and 650 days post-implant showed sustained satisfaction, with independence rated 5/5 during use and 4/5 during setup. T19, who cannot use eye tracking reliably and types via attempted NATO codeword speech timed to ventilator exhalations, was successfully onboarded onto the same finite-state-machine backend with a redesigned timed-selection frontend, demonstrating the architecture's adaptability across very different motor and speech residuals.
Relevance
For accessibility practitioners, this paper is one of the first detailed accounts of BCI use as a daily, sustained AAC and computer-control modality, not as a lab demonstration. The headline lessons translate beyond BCI: control-mode redundancy (eye tracking plus neural cursor) preserves independence when one modality fails; user-initiated recalibration is essential because dependence on lab staff defeats home use; and a shared finite-state-machine backend with swappable frontends dramatically reduces the cost of personalizing for the next user. The team explicitly maps their work onto Wobbrock and colleagues' ability-based design principles and meets four of seven (ability, accountability, adaptation, transparency), with performance, context, and commodity left as open challenges - a useful self-audit format others could borrow. Limitations to flag: this work covers only two participants enrolled in a clinical trial of a surgically implanted device, so generalizability and accessibility outside research settings remain limited until BCI commercialization matures. Even so, the practical design playbook - co-design over 22 months, technology-probe deployment, longitudinal surveys, and a shared adaptable architecture - is directly transferable to other emerging assistive technologies.
Tags: brain-computer interface · AAC · ALS · amyotrophic lateral sclerosis · paralysis · speech neuroprosthesis · co-design · personalization · ability-based design · longitudinal deployment