Responsive neurostimulation

Most of the time, even in severe epilepsy, the cortex at a seizure focus is not seizing. A drug acts all day regardless, and so does a continuous stimulator. Responsive neurostimulation acts only when asked: an implant reads the electrocorticogram at the focus, runs a detector on it, and when the detector fires delivers a brief burst of current through the same leads.

The premise had support from the clinic. Mapping cortex through subdural electrodes sometimes provokes afterdischarges, runs of epileptiform activity that outlast the stimulus, and in 1999 Lesser and colleagues found that about half of them stopped within two seconds of a brief burst of pulses, against fewer than one in twenty left alone. NeuroPace’s RNS System was approved by the FDA in November 2013 for adults whose focal seizures come from at most two foci and have resisted at least two drugs — closed-loop brain stimulation, more than eleven years before deep brain stimulation gained an adaptive mode.

Three cheap detectors

The implant is a curved titanium can set flush into an opening cut in the skull, under the scalp, wired to one or two leads of four contacts each — strips on the cortex or depth electrodes in it. It records four channels, each the difference between neighbouring contacts, at 250 samples a second, and runs its detectors on them continuously, for years, on a battery that cannot be recharged.

All three detectors are cheap to compute. Line length sums the absolute differences between successive samples over a window, so a single number grows with amplitude and with frequency, for a subtraction and an addition per sample. Area tracks signal power. The third, listed as bandpass and also called a half-wave tool, looks for rhythm in a programmed band: a half wave, the swing from one turning point of the signal to the next, lasts half a period, so a window on its duration is a window on frequency. In the pivotal trial 98% of patients used one, and the commonest setting, 0.5–125 Hz, runs right up to the Nyquist limit of the sampling.

A detection triggers short trains of current-controlled, charge-balanced biphasic pulses — by default 200 Hz, 160 µs pulses, 100 ms bursts. It is a small dose. Over nine years of follow-up, patients averaged 1,028 detections a day, most answered by a single therapy of two 100 ms bursts: 3.4 minutes of stimulation a day on average and 10.3 at most, about a quarter of one per cent of the time. Thalamic stimulation for epilepsy, approved in 2018, ran one minute in six in its pivotal trial.

A loop with a clinician in it

As a feedback controller the implant is spare: a detector, a threshold, a preset burst and a daily cap on therapies, all set from outside. The FDA’s summary says plainly that it “is not a seizure detection device”; it answers patterns a physician has picked out of the patient’s own recordings.

So the physician is the outer, adaptive loop. Patients read out the implant at home with a wand — over an analogue telephone line, in 2013 — and the physician reviews what it stored and reprograms it. In the pivotal trial, initial detection settings were changed for 83% of patients after their stored recordings were reviewed, and stimulation settings for all but 4 of 191 within two years.

What the physician can review is rationed by memory. The implant holds about 30 channel-minutes of electrocorticogram — seven and a half minutes of all four channels, typically as 90 s records triggered by a long detection, a schedule, a saturated amplifier or the patient’s magnet — and overwrites the oldest. Beyond that it keeps counters, among them hourly detection counts for the past 28 days. It cannot keep a complete record, so the trials measured efficacy the old way, from patient diaries.

Multi-day cycles of seizure risk

The counters had a second use. An hourly count of detections can be read as an hourly measure of interictal epileptiform activity, the discharges an EEG shows between seizures, and as long as patients uploaded at least every four weeks it ran unbroken for years — from inside the head, through ordinary life.

In 2018 Maxime Baud, Vikram Rao and colleagues at UCSF analysed those counts in 37 patients, over 3 months to 9.9 years each. The activity oscillated with the time of day, and also on multi-day, or multidien, cycles particular to each patient, most often 20–30 days long and stable for up to a decade. Seizures arrived preferentially on the rising phase of the multidien cycle.

Cycles in seizure timing were not new — Griffiths and Fox reported rhythm in epilepsy in 1938, and Karoly and colleagues found weekly and longer cycles in seizure diaries the same year as Baud — and they are not confined to women, so the menstrual cycle cannot explain them. What the implant added was the cycle in the brain’s own activity between seizures, with seizures clustered on one phase of it. Across the trial cohort, Leguia and colleagues found multidien seizure cycles in 60% of those who could be assessed, and Proix and colleagues forecast next-day seizure risk from the same counters, better than chance for two-thirds of a validation cohort (both 2021).

That is what crossed the border: seizure risk as a slow, patient-specific oscillation that took years of intracranial recording to see, measured by a therapy that was logging so that clinicians could tune it.

Why it improves for years

In the pivotal trial’s 12-week blinded phase, among 191 adults, seizures fell 37.9% with stimulation and 17.3% with sham. The median reduction then rose to 44% at one year and 53% at two, and to 75% at nine, by which time 18.4% of the 256 people implanted had gone at least a year without a seizure. Everything after the blinded weeks was open-label, though the nine-year study’s alternative analyses suggest the rise is not merely a matter of who stayed in.

A stimulus that ended seizures one at a time should not need nine years to get good, and as Khambhati and colleagues put it, stimulations greatly outnumber seizures and fall mostly between them. In stored recordings from 11 patients, Kokkinos and colleagues found outcome unrelated to the direct effects of stimulation on detected seizures but related to changes remote from it (2019). Khambhati’s group found that patients who did best showed a progressive, frequency-dependent reorganisation of connectivity between seizures, emerging within a year and scaling with seizure reduction (2021). The honest summary is that a device built to answer each seizure may work mainly as slow neuromodulation gated by the brain’s own activity — a plant that changes because of its controller.

What it costs

It is surgery, and repeated surgery. Serious infection at the implant site reached 12.1% of patients over the whole follow-up, at 4.1% per procedure, and the first model’s battery lasted a median of about three and a half years before another procedure replaced it; a later model, the RNS-320, is expected to reach eight at moderate use. Some patients chose not to have it replaced.

The 250 Hz sampling puts everything above 125 Hz out of sight. And tuning remains labour-intensive and empirical: detectors are biased towards sensitivity, and most detections are never checked against a recording.