Sub-Scalp AI Chip Targets Days-Long EEG Monitoring for Epilepsy

An investigational sub-scalp EEG chip designed at Northeastern University could extend continuous brain monitoring from the 20 to 90 minutes of a standard clinical EEG to days or months, while running on-device machine learning to flag seizure signatures in real time.

The project, led by electrical and computer engineering professor Aatmesh Shrivastava with funding from the National Institutes of Health, places electrodes beneath the skin rather than on its surface. Nerve cells fire up to 1,000 times per second, and the closer-to-source placement reduces the overlap that occurs when large populations of neurons discharge simultaneously through bone and scalp.

"By going a little bit below, you can actually improve the quality of the signal," Shrivastava said.

What can continuous sub-scalp monitoring reveal that routine EEG misses?

Routine scalp EEG captures brief snapshots of cortical activity. Long-duration recording exposes patterns a 30-minute study cannot, Ziv Peremen, CEO of X-trodes Ltd, said. It can show "transitions into and out of abnormal activity, relationships to sleep and behavior, and early signatures that may precede clinical events."

The timeline matters clinically. Manfred Hartmann, senior research engineer at the AIT Center for Health & Bioresources in Vienna, noted that "seizures … often follow individual rhythms over days to months." Multi-day recordings let clinicians verify whether a patient is "truly seizure-free" and estimate windows of elevated risk, he added.

Ivan Gligorijević, CEO of wearable EEG firm mbraintrain.com, said chronic monitoring offers a more nuanced view of medication effects and brain function than episodic testing.

How does the chip turn raw EEG into actionable markers?

The on-chip machine learning decomposes each recording into signal attributes — amplitude, frequency and waveform shape — then links those features to clinical events logged during acquisition. The network stores combinations of attributes as candidate seizure markers and refines its profile as more data arrives.

Shrivastava compared the platform to a continuous glucose monitor. "It is not an intervention," he said. "You can analyze the data, identify disease patterns, identify … aspects of life that you can control."

Geoff Bobb, executive director of Epilepsy Toronto, stressed that seizure morphology varies by region of onset. "Seizures in different parts of the brain will manifest differently," he said, and convulsions, sensory phenomena and amnestic events call for distinct treatment approaches.

Where does the device sit in the current monitoring landscape?

Existing continuous sub-scalp systems act largely as sensors, Shrivastava noted, relaying raw EEG for offline review by a clinical neurophysiologist. The Northeastern chip embeds inference at the recording site, flagging events in real time rather than retrospectively. That distinction could reduce reading-room workload if the markers hold up across patient cohorts.

Power budget drives the design. Shrivastava's earlier work on low-power and self-powering chips — work that led to his induction into the National Inventors Academy — underpins the energy strategy for implanted operation. Continuous use requires energy harvesting, since an implanted device cannot be recharged nightly.

What happens next?

Once bench performance meets spec, the prototype will undergo animal testing in rats or macaques at the University of Florida, according to the source. Human feasibility studies and any subsequent FDA review would follow.

Tracy Dixon-Salazar, president and CEO of the Lennox-Gastaut Syndrome Foundation, called the project "the kind of innovation our community has been waiting for." If the preclinical work supports the design, Shrivastava's team will move toward first-in-human studies of a chronic, on-chip-analysis EEG platform intended for outpatient monitoring of epilepsy and, eventually, other conditions involving disrupted cortical signaling.