BRENNEN

Industries / Frontier science / Brain-computer interfaces

Stimulating the brain

Stimulation sends small electric pulses into the brain through implanted electrodes, to calm a tremor, stop a seizure, or create a sense of touch or sight.

7

Milestones achieved, from 1968 to 2025.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1968

    Points of light

    Giles Brindley and Walpole Lewin implanted radio receivers wired to electrodes on the visual cortex of a 52-year-old blind patient. Radio signals made the patient see phosphenes, small spots of light, at fixed places in the field of view. Electrodes 2.4 mm apart gave spots that could be told apart, and stimulating several at once produced simple patterns.

    ebi.ac.uk

    Credit

    G. S. Brindley and W. S. Lewin, in The Journal of Physiology

  2. 1987

    Steady pulses stop tremor

    During surgery that destroys a small spot in the thalamus to treat tremor, Alim-Louis Benabid's team saw that stimulating the same spot 100 times a second stopped the tremor. For patients with tremor on both sides, they destroyed the spot on one side and implanted stimulating electrodes on the other; stimulation strongly reduced tremor, though less completely than the lesion, partly because the stimulator could not go above 130 Hz. On 31 July 1997 the US Food and Drug Administration (FDA) approved Medtronic's Activa Tremor Control System for this deep brain stimulation (DBS).

    ebi.ac.ukaccessdata.fda.gov

    Credit

    A. L. Benabid and colleagues; Medtronic

  3. 2013

    A stimulator that listens first

    NeuroPace's RNS System, approved by the FDA on 14 November 2013 for adults whose focal epilepsy drugs do not control, records brain activity and sends short pulses only when it detects patterns set for each patient. In the blinded period of its main trial, seizures fell by an estimated 37.9% with stimulation against 17.3% with sham, where the device was implanted but left off, according to the FDA's summary.

    accessdata.fda.govaccessdata.fda.gov

    Credit

    NeuroPace

  4. 2017

    Depression trial falls short

    Stimulation of the subcallosal cingulate, a region below the front of the brain's midline, had shown promise for depression that other treatments had not helped. BROADEN, a trial that compared it with sham stimulation, was halted at a planned futility check: 20% of 60 people given stimulation responded, against 17% of 30 given sham.

    Lessongains seen when everyone knows the device is on can disappear against a sham control.

    ebi.ac.uk

    Credit

    P. E. Holtzheimer and colleagues, in The Lancet Psychiatry

  5. 2021

    Touch speeds a robot hand

    In 2016 a University of Pittsburgh team showed that intracortical microstimulation, tiny currents through electrodes in the touch area of the cortex, made a man with a spinal cord injury feel pressure at particular spots on his hand, stable for months. In 2021 they fed signals from a robot hand's sensors back this way while he moved the arm with his motor cortex. His median time on a standard arm test fell from 20.9 to 10.2 seconds.

    ebi.ac.ukpmc.ncbi.nlm.nih.gov

    Credit

    S. N. Flesher, R. A. Gaunt, J. L. Collinger and colleagues, in Science Translational Medicine and Science

  6. 2021

    Shapes seen through an array

    A 57-year-old blind person had a 96-electrode array in the visual cortex for six months. Stimulating several electrodes at once lowered the current needed and produced spots of light the person could tell apart, enough to identify some letters and the edges of objects. Implanting and removing the array caused no complications.

    ebi.ac.uk

    Credit

    E. Fernández, R. A. Normann and colleagues (Miguel Hernández University, Spain, and University of Utah), in the Journal of Clinical Investigation

  7. 2025

    Stimulation that adjusts itself

    The FDA approved Medtronic's BrainSense Adaptive DBS for Parkinson's disease in February 2025. Standard DBS runs at fixed settings; this implant senses brain activity and changes stimulation in real time. In the ADAPT-PD trial, published in JAMA Neurology, 45 people tried it for 30 days and 40 used it for over a year. In its dual-threshold mode, people reported 1.3 more hours a day of good control and 1.6 fewer hours undertreated than on their usual settings, in an unblinded comparison, and 44 of 45 chose to keep it.

    healio.comnews.medtronic.com

    Credit

    Medtronic; H. Bronte-Stewart (Stanford University), the trial's global principal investigator

Watch for

  1. Six years of Orion

    At the North American Neuromodulation Society meeting in January 2026, Nader Pouratian of UT Southwestern Medical Center presented six-year results of Orion, a visual cortex implant that began at Second Sight. Six blind people were implanted in 2018 and 2019; all devices still worked, fewer than 4% of electrodes had stopped working, and there was one serious adverse event, a seizure early in the study. Cortigent, which now develops Orion, plans a larger pivotal trial with the FDA.

    Achieved whenthe six-year results, with functional test scores, appear in a peer-reviewed journal.

    neurotechreports.com

    Claimed by

    the study's investigators and Cortigent, January 2026

  2. A wireless implant for sight

    Illinois Institute of Technology says its Intracortical Visual Prosthesis, which it calls the first visual cortex implant built from a group of small, fully implanted wireless stimulators, went into its third participant in May 2026: 34 stimulators carrying 544 electrodes. It says four years of testing since the first implant show improved ability to move around and to do basic tasks guided by sight.

    Achieved whenmeasured navigation and task scores for the participants are published in a peer-reviewed journal.

    iit.edu

    Claimed by

    Illinois Institute of Technology, May 2026

Open

  1. 01

    One test for vision encoders

    A public benchmark for encoders, the software that turns a camera image into a stimulation pattern for a visual implant. Each encoder's output is passed through one fixed, published simulator of cortical phosphenes, and the simulated views are scored automatically by off-the-shelf image recognition models on a fixed set of everyday scenes (objects, letters, doorways), with a leaderboard anyone can submit to. No person views anything and no device touches anyone.

    The brief
    Credit · unclaimed
  2. 02

    Threshold rules under test

    Add the simple rules behind approved adaptive DBS, on-off, single-threshold and dual-threshold control, which raise or lower stimulation when a beta-band brain rhythm crosses set levels, to DBS-Gym, an open simulator of Parkinsonian brain rhythms. Then publish a head-to-head with its continuous, PID (proportional-integral-derivative) and reinforcement-learning controllers across many simulated patients with drift and noise.

    The brief
    Credit · unclaimed

Open 01 · the brief

One test for vision encoders

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public benchmark for encoders, the software that turns a camera image into a stimulation pattern for a visual implant. Each encoder's output is passed through one fixed, published simulator of cortical phosphenes, and the simulated views are scored automatically by off-the-shelf image recognition models on a fixed set of everyday scenes (objects, letters, doorways), with a leaderboard anyone can submit to. No person views anything and no device touches anyone.

Why it is still open

Encoders are compared paper by paper, each with its own test. The dynaphos simulator paper (2024) optimises encoders inside its own pipeline, and the human-in-the-loop work of Jacob Granley, Michael Beyeler and colleagues scores encoders with its own digit classifier and sighted volunteers. No shared scene set or leaderboard was found.

ebi.ac.ukarxiv.org

Built with what exists

dynaphos, an open PyTorch simulator of phosphenes from visual cortex implants that runs in real time (M. van der Grinten and colleagues, eLife 2024); open image recognition models; PyTorch; a hosted leaderboard on EvalAI, which the Neural Latents Benchmark for decoders already uses. Simulators that can be differentiated, so encoders can be trained through them, and capable open vision models are the advances that let a computer act as the first judge.

arxiv.org

  1. First step

    Install dynaphos, render 100 everyday photos through its default settings with a plain edge-detection encoder, and score them with an open object classifier.

  2. Done looks like

    At least 1,000 test images across 5 or more scene types; at least 6 published encoders scored at two implant sizes (for example 100 and 1,000 phosphenes); every score reproducible from released code.

Open 02 · the brief

Threshold rules under test

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

Add the simple rules behind approved adaptive DBS, on-off, single-threshold and dual-threshold control, which raise or lower stimulation when a beta-band brain rhythm crosses set levels, to DBS-Gym, an open simulator of Parkinsonian brain rhythms. Then publish a head-to-head with its continuous, PID (proportional-integral-derivative) and reinforcement-learning controllers across many simulated patients with drift and noise.

Why it is still open

DBS-Gym, published at KDD 2025, ships baseline controllers for continuous high-frequency DBS and PID control plus tested reinforcement-learning methods, but no threshold rules. Fleming, Dunn and Lowery's 2020 model tested on-off and dual-threshold controllers against P and PI control, without learning-based controllers or the drift and noise DBS-Gym adds. A January 2026 preprint on bandit algorithms, a light kind of learning controller, compared them with deep reinforcement learning in a computer model of the basal ganglia and measured their energy on microcontrollers, but its abstract names no threshold rule among the comparisons. No study comparing all of these on one benchmark was found.

github.comarxiv.orgpmc.ncbi.nlm.nih.govarxiv.org

Built with what exists

DBS-Gym (Python, MIT licence, a gymnasium environment); standard reinforcement-learning libraries; the dual-threshold rule as defined by Fleming and colleagues (a target range between the 10th and 20th percentiles of beta activity with DBS off); the ADAPT-PD trial's published statistical plan, which names the single- and dual-threshold modes.

cdn.clinicaltrials.gov

  1. First step

    Clone DBS-Gym, run its PID baseline in the simplest environment, and add a dual-threshold controller using Fleming's rule.

  2. Done looks like

    At least six controllers run in all three of DBS-Gym's preset environments with at least 20 random seeds each, scored on beta suppression and total energy delivered, with code and results public.