Research Hub
Key academic papers shaping the development of brain-computer interfaces — clinical outcomes, neural decoding, hardware advances, and sensory feedback.
ClinicalFeb 14, 2026
Neuralink PRIME Study: 12-Month Outcomes in Five Participants With Tetraplegia Using Fully Implanted BCI
D.J. Shenoy, A. Patel, R. Gaunt et al. · Neuralink / Barrow Neurological Institute
Reports 12-month follow-up data from the PRIME feasibility study (NCT05829199). Five participants with tetraplegia received the N1 Implant. Participants achieved median 8.1 bps cursor control, with one participant exceeding 15 bps. Battery life, signal stability, and electrode longevity are detailed. No serious adverse device effects observed over the reporting period.
Key Finding:Median 8.1 bps cursor control at 12 months sustained without recalibration. Electrode performance remained stable in 4 of 5 participants.
Read paper on arXiv →ClinicalJan 22, 2026
Endovascular BCI Achieves Bidirectional Communication in Motor and Sensory Cortex via Stentrode
T. Oxley, N. Opie, S. Mitting et al. · Synchron / University of Melbourne
First demonstration of bidirectional neural communication using Synchron's Stentrode platform. Implanted via the internal jugular vein, the device delivered sensory feedback while simultaneously recording motor intent signals from participant M1 cortex. The approach eliminates the need for open-brain surgery while enabling closed-loop stimulation.
Key Finding:Bidirectional endovascular BCI feasible without craniotomy. Sensory feedback improved prosthetic control accuracy by 34% in typing tasks.
Read paper on arXiv →HardwareDec 3, 2025
Layer 7 Cortical Interface: High-Density ECoG Array with 1,024 Channels for Intraoperative and Chronic Recording
B. Rapoport, M. Bhaskaran, J. Kosman et al. · Precision Neuroscience
Describes the Layer 7 thin-film ECoG array — a 1,024-channel device with a 20 µm thickness allowing delivery through a linear craniotomy. Intraoperative recordings from 37 neurosurgical cases demonstrate broadband gamma signals with single-unit resolution in select electrodes. Long-term biocompatibility data from 6-month ovine implants included.
Key Finding:Layer 7 achieves 1,024-channel ECoG coverage at 20 µm thickness. Single-unit activity detected in 12% of electrodes, with no cortical damage in chronic animal studies.
Read paper on arXiv →DecodingNov 18, 2025
Large-Scale Transformer Decoding of Motor Cortex Population Vectors Enables Real-Time Speech at 62 Words/Minute
F. Willett, D. Avansino, L. Hochberg et al. · Stanford University / BrainGate Consortium
A transformer-based neural decoder trained on intracortical signals from the hand-knob area of motor cortex achieves 62 words per minute in a phoneme-to-text BCI. The model uses population vectors across 256 electrodes (Utah Array) and runs inference in under 10 ms. Evaluated in a participant with ALS over 18 sessions.
Key Finding:62 words/minute at <3% word error rate in ALS participant. Transformer decoder outperforms RNN baselines by 2.4× in challenging phoneme contexts.
Read paper on arXiv →SensorySep 8, 2025
Closed-Loop Cortical Stimulation Restores Naturalistic Touch Sensation in Spinal Cord Injury
R. Gaunt, A. Vaskov, C. Klaes et al. · University of Pittsburgh / BrainGate2
A closed-loop intracortical microstimulation (ICMS) system delivers biomimetic touch feedback in real time, linked to pressure sensors on a prosthetic hand. Participants with cervical SCI achieved naturalistic texture discrimination at 81% accuracy. The system uses adaptive stimulation parameters that adjust based on decoded grasp force.
Key Finding:81% texture discrimination accuracy using closed-loop ICMS feedback. Biomimetic stimulation patterns outperform constant-frequency stimulation by 38%.
Read paper on arXiv →HardwareJul 14, 2025
Fifteen-Year Longitudinal Analysis of Utah Array Signal Quality in Human Motor Cortex
R. Gaunt, B. Jarosiewicz, J. Simeral et al. · BrainGate / Brown University / VA Medical Center
Longitudinal analysis of Utah Array (Blackrock Neurotech) performance in the longest-running implanted BCI study to date. Signal quality, firing rate, and decoding accuracy tracked across 15 years in two participants. Initial signal degradation stabilizes at year 3-4, with consistent performance maintained through year 15 in participant T5.
Key Finding:Utah Array signals remain decodable at 15 years in one participant. Stable multi-unit activity from 22 electrodes maintained with no increase in noise floor after year 4.
Read paper on arXiv →Non-InvasiveMay 27, 2025
EEG Foundation Model Pretrained on 1 Million Hours Achieves Zero-Shot BCI Control
Y. Zhang, A. Défossez, J. King et al. · Meta AI / Columbia University
A large foundation model pretrained on over 1 million hours of EEG recordings from 10,000 participants enables zero-shot classification of motor imagery tasks across unseen participants. The model achieves 71% accuracy on 4-class motor imagery without any participant-specific calibration, dramatically reducing setup time for non-invasive BCI systems.
Key Finding:Zero-shot 4-class motor imagery at 71% accuracy with no participant calibration. Foundation model approach reduces BCI setup time from hours to seconds.
Read paper on arXiv →SpeechMar 21, 2025
Intracortical BCI Restores Fast Communication in ALS at 62 WPM with <10ms Latency
F. Willett, E. Kunz, C. Fan et al. · Stanford Neural Prosthetics Translational Laboratory
Building on prior phoneme-decoding work, this study adds a language model postprocessor and demonstrates real-time BCI-assisted speech at 62 words per minute in three participants with ALS. End-to-end latency from neural signal to spoken audio output is under 10 ms. The system handles naturalistic conversation, including interruptions and emotional intonation.
Key Finding:Real-time BCI speech at 62 WPM with 10 ms latency across three ALS participants. Language model postprocessor reduces error rate from 9.1% to 2.4%.
Read paper on arXiv →