Most neurotechnology listens in one place. An implant sits in a region, records there, and when something looks wrong it pushes current back. That model strains when a psychiatric illness is understood not as a defect in a single area but as a disruption in the conversation between several.

Researchers at EPFL have now built a chip that takes that idea literally. LogSync, developed in Mahsa Shoaran's Integrated Neurotechnologies Lab, is an implantable system-on-chip about 4 × 2 millimetres in size that watches not one region but the relationship between regions.

What the chip measures

LogSync records brain activity and extracts two biomarker-like quantities from it: the strength of oscillations — how much energy sits in given frequency bands — and the synchrony between signals from different brain regions. A machine-learning model then works on those network-level features rather than on raw signal, mapping them onto cognitive or emotional states. When a state is recognised, the chip can issue a stimulation trigger.

The difference from ordinary deep brain stimulation is categorical. Conventional systems run open-loop: they deliver pre-set pulses regardless of what the brain is doing. Even closed-loop, adaptive systems usually hang off a single recording site. For psychiatric conditions that is a structural weakness, because the clinically relevant information is distributed across a network instead of sitting on one contact.

The real trick is doing the maths in logarithms

An implantable chip has a hard budget, and it is not compute that blows it — it is current draw. The team describes two decisions that had to fit together.

First, hardware and algorithm were co-designed, with the implant's limits as the starting point rather than an afterthought. Second, LogSync computes in the logarithmic domain. The multiplications and divisions needed to analyse synchrony and oscillations become additions and subtractions, which are far cheaper in silicon.

The result, according to the researchers, is under 200 microwatts for the whole system — a fraction of what many implantable neural interfaces require. The 2025 precursor, a 32-channel design, came in at 196 microwatts, which gives a sense of how tight the arithmetic is here.

What was actually demonstrated

The validation is careful but indirect, and that matters for how the result should be read.

For decoding, the researchers used human brain recordings from earlier psychiatric experiments: one dataset capturing shifts in cognitive engagement during a mental task, another capturing emotional arousal as participants watched video clips. From the neural signals alone, the algorithm distinguished those states. That is an analysis of recorded data, not a real-time measurement inside a patient's head.

The closed loop was tested in rats at the University of Minnesota, where the implant tracked ongoing oscillations and delivered stimulation at pre-defined phases of brain activity. Phase-locked triggering is the technically awkward part, because it requires the circuit to recognise a state fast enough to act inside the right window.

So what the chip establishes is that network-level biomarkers are computable on an implant at this power budget, and that precisely timed triggers can be derived from them. What it does not establish is that such control helps patients. That needs clinical trials, and there are none yet.

Why this is more than a circuit detail

The work appears in the IEEE Journal of Solid-State Circuits and grew out of a CHF 6.6 million grant from the US National Institute of Mental Health, which is funding implantable systems to measure and control oscillatory brain networks in severe psychiatric illness — mood and anxiety disorders above all.

What is notable is the direction the lab is taking in parallel: away from hand-defined biomarkers and towards models that learn directly from recordings which patterns matter across regions, channels and time. Their ICML 2026 paper, RECTOR, uses self-supervised learning for exactly that. A spin-off, Infera Neuro, is meant to carry the results towards clinical use.

If psychiatric illness is a network disorder, the question stops being which electrode to listen to and becomes how to follow a conversation. A chip 4 × 2 millimetres across, running on under 200 microwatts, is a serious start — but only the start of the evidence chain.

Sources