Brain-Computer Interfaces in Numbers: Trials, Implants and Research

A brain-computer interface (BCI) records brain activity and turns it into commands: a cursor moves, a robotic arm grasps, a synthetic voice speaks words a person can no longer say. AI models do the decoding, translating patterns of electrical activity into intended movements, letters or sounds. To see how far the field really is, we analyzed all 280 studies on ClinicalTrials.gov, the largest public registry of clinical studies, whose entry mentions a brain-computer or brain-machine interface, and counted the research papers on PubMed.

Two kinds of brain-computer interfaces

Non-invasive systems read brain activity from outside the head, mostly with an EEG cap of electrodes on the scalp, sometimes with near-infrared light (fNIRS). They are safe and cheap, but the skull blurs the signal, so they work for simple choices such as “left or right” rather than for typing at speed. Implanted systems place electrodes on or in the brain: tiny needle arrays in the motor cortex, thin grids on its surface (ECoG), or electrodes inside a blood vessel next to the brain. They need surgery, and in return the signal is clear enough to control a cursor, a robotic arm or synthetic speech.

The difference shows in the size of the studies. Half of the implant studies enroll or plan to enroll 6 participants or fewer, half of the non-invasive studies 30 or fewer. In the US, a trial with an implanted device that carries a significant risk needs an investigational device exemption from the FDA, and every participant needs brain surgery, so implant studies start with a handful of people.

Who the studies are for

Stroke is the most common condition, with 112 studies. Most of them are rehabilitation trials: a patient imagines moving a paralyzed hand, an EEG-based BCI detects the attempt and triggers a robotic glove or electrical muscle stimulation, which links intention and movement again. The implant studies focus on other patients: 40 of the 59 concern a spinal cord injury or paralysis of all four limbs, 26 concern ALS, a disease in which people gradually lose control of all muscles, including those needed for speech.

Where the research happens

Studies with sites in the United States lead by far. The studies cover 35 countries in total, and only 13.2% are led by companies; most come from universities and hospitals.

The implant trials running now

Implants are the fastest-growing part of the registry: 35 of the 59 implant studies started in 2024 or later, and 46 are recruiting or still running. 18 of these ongoing studies have sites in the US and 11 in China. The titles show where the field is heading: wireless systems that are fully implanted instead of a cable through the skull, more recording channels, and speech decoding, including a study of speech in Mandarin Chinese.

Research output

PubMed lists 1,390 papers from 2,025 with a brain-computer or brain-machine interface in the title or abstract, 2.4 times as many as in 2015. Research grows much faster than clinical use: in the same year, only 36 new studies with participants started.

What the numbers cannot tell

A registry entry shows what a team plans, not what it achieved; results appear later, if at all. Not every study is registered at ClinicalTrials.gov: studies that only appear in other registries, such as the Chinese Clinical Trial Registry, are missing here. The split between implanted and non-invasive studies is our own, based on keywords in each entry, and 92 entries do not say clearly enough which kind of interface they use.

How we counted

We retrieved all studies from the ClinicalTrials.gov API whose entry contains “brain-computer interface” or “brain-machine interface”, with or without a hyphen, and grouped them by start date, status, condition, country of the study sites and type of interface. Implanted means the entry mentions an implant, intracortical or surface electrodes on the brain, depth electrodes or a device in a blood vessel; non-invasive means it mentions EEG, fNIRS, MEG or a non-invasive method. Papers were counted on PubMed by publication year with the same terms in the title or abstract. The source and date are below each table.

How such an interface turns brain signals into commands is explained in our learning path in Physical AI.