Research

Sleep is read in five stages — wakefulness, REM, and the three depths of non-REM — and it is scored one thirty-second epoch at a time. The convention was built so that a technologist could mark up a night’s chart by hand, and it has served clinical sleep medicine for decades.

It also fixes three things in advance: how many states there are, how briefly one may last, and that the brain moves between them on a tick of the clock. None of the three is a finding. They are working assumptions, and a continuous intracranial recording is in an unusual position to test them.

So the question here is what a scored stage stands in for. Whether a transition is a border the brain crosses or a slope it descends. Whether a measurement means the same thing in sleep as in wakefulness — not safe to assume, and examined directly for SEEG connectivity across recording conditions and medication states. And whether the states that matter clinically are the five on the chart at all, given that network signatures define consciousness state during focal seizures.

Stereo-electroencephalography places fine depth electrodes inside the brain and records continuously for days, because a person with drug-resistant focal epilepsy needs the tissue that starts their seizures identified before surgery can help them. The recording exists for their care. What it gives in addition is the closest sustained view anyone has of an intact human brain doing ordinary things — asleep, awake, and everything in between.

Most of those days are not seizure. They are the long stretch in between, and a line of work running through this group’s papers argues that the quiet is not empty. Regions that generate seizures show high inward-directed connectivity at rest: they look less like tissue that is unusually excitable and more like tissue the rest of the network is holding down. That suppression is the dominant differentiator of seizure onset zones in focal epilepsy — an idea first set out with network-level supporting evidence — and a seizure spreads when the suppressive network collapses.

If that account holds, the informative measurement is made between events rather than during them. Which is fortunate, because between is nearly all of the record.

A state is easier to name than to locate. The definition we work from is geometric: a region of a space whose coordinates are learned from the recording rather than chosen in advance, with a transition as the path that leaves one region for another. Candidate coordinates are themselves testable — the time-resolved correlation of distributed activity, for one, tracks excitation–inhibition balance.

The definition is only worth holding if the geometry is shared. A model that must be retrained on every new brain has described that brain, not brain states; KenazLBM is where we are testing whether one learned space generalises instead of assuming it does. The reason to care is a device. Adaptive neuromodulation has to estimate the state it is responding to, in the moment, on a person it was not trained on — the case for doing that from SEEG, and an earlier closed-loop formulation.