Some brain circuits can sit in either of two stable states under the same conditions: a quiet resting state or a self-sustaining rhythm. A seizure is the pathological version of the rhythm. Pushing such a circuit back to rest with a brief electrical pulse works only if the pulse lands at the right moment in the rhythm. The identical pulse can end a seizure or prolong it, and the window that separates the two outcomes is tens of milliseconds wide. Current control theory for networks of these circuits assumes smooth, continuous feedback on a fixed wiring diagram. Real stimulators deliver discrete pulses, the right timing is unknown in advance, and the coupling between brain regions keeps changing.
We propose a way to find the right pulse timing online, cheaply, and in a way that follows the circuit as it changes. A small probe pulse shifts the timing of the next few cycles of the rhythm, and that shift is easy to measure. Work in Parkinsonian rats found that how much a pulse shrinks the rhythm is well predicted by the negative slope of how much it shifts the rhythm's timing. If that relationship also holds for seizure-like circuits with two stable states, a stimulator can learn from harmless timing probes where the "quench window" lies. It can re-learn the window as the balance between excitation and inhibition drifts, and it can choose which few regions to stimulate from the current coupling pattern rather than a fixed one.
Core Claim
In a network of bistable excitatory–inhibitory nodes, brief pulses delivered to a small, re-selected subset of nodes can return the whole network from a seizure-like limit cycle to rest. A limit cycle here means a stable repeating oscillation. This works with less delivered charge and less risk of prolonging the seizure than fixed-phase or continuous-feedback control, provided each pulse is timed to the phase where the amplitude response is most negative. That phase can be estimated from phase-response probes through the relation
$$\mathrm{ARC}(\theta) \approx -c\,\frac{d\,\mathrm{PRC}(\theta)}{d\theta},$$
where $\theta$ is the phase of the ongoing rhythm at pulse delivery, $\mathrm{PRC}(\theta)$ is the phase shift the pulse causes, $\mathrm{ARC}(\theta)$ is the fractional change in oscillation amplitude it causes, and $c>0$ is a scale factor we fit per node. The relation comes from phase-reduction theory of weakly perturbed oscillators. Mirkhani et al. confirmed it empirically in 13 Parkinsonian rats during pallidal stimulation, with a mean correlation of $R=-0.84$.
Nobody has tested it for bistable nodes, where the stable oscillation coexists with a stable rest state and the pulse must carry the state across the boundary between them. Testing this is the first thing the project does. If the relation fails there, the rest of the method loses its cheap estimator, and we would report that.
How This Sits Against Prior Work
The closest work is Liu et al.. It models seizures as coupled bistable nodes, with rest at the origin and a seizure limit cycle, and proves that continuous state feedback $u_i=-q_ix_i$ on a few "driver" nodes quenches the whole network. In their 8-node example, 3 drivers suffice. Their guarantee depends on the second-smallest eigenvalue of a fixed, undirected graph Laplacian, which measures how tightly the graph is connected. It also assumes linear diffusive coupling. We keep their pinning idea, meaning control of a subset of nodes, but replace continuous feedback with phase-timed pulses. We also make the coupling nonlinear, directed, and time-varying.
Taylor et al. motivate the timing problem. In a thalamocortical model, a pulse 0.15 s after seizure onset terminated spike-wave discharges, while the same pulse at 0.20 s prolonged them. That result covers a single circuit with a hand-found time, and we want the window found automatically across a network.
Mirkhani et al. supply the estimator. They treat the target as one oscillator population and characterize its responses; they do not use them to quench a bistable network. Shin et al. constrain the design from the hardware side. Their 60 µW chip extracts phase with a lightweight interpolator that correlates above 95% with an ideal Hilbert transform. It compensates filter delay but stimulates at a fixed target phase (0° of 6 Hz theta). Our method must fit within that kind of phase-estimation budget, and it asks when a fixed target is not enough.
The excitation–inhibition literature matters here because it identifies the variable that moves the quench window. Normative balanced networks (Boerlin et al.; Efficient coding in biophysically realistic excitatory-inhibitory spiking networks) and the fast "balanced amplification" regime seen in primate area MT (Pattadkal et al.) show how tight inhibition prevents latching into persistent states. None of them models the path by which weakening inhibition produces a seizure limit cycle. We treat inhibitory gain as the slow drift parameter that reshapes the response curves.
Work on dynamic connectivity supplies the evidence that coupling changes before seizures. MB-dMGC-CWTFFNet learns time-varying graphs that predict seizures on CHB-MIT, and STAGIN shows gains from sliding-window graphs over static ones. Both only decode; neither controls. BRep is the counterweight: a good static representation beat dynamic graph models on ADHD-200. We therefore treat dynamic driver selection as a hypothesis to be ablated, not a given.
Research Program
The model. Each node is a Wilson–Cowan excitatory–inhibitory pair, a two-population firing-rate model. Its parameters are chosen in the regime where a low-activity fixed point and a large limit cycle coexist, and inhibitory gain $g$ is the parameter that moves the node toward or away from the bifurcation that destroys the rest state. Nodes couple through directed excitatory connections $W(t)$. We use two versions of $W(t)$. The first is held fixed, which matches Liu's setting. The second follows trajectories of time-varying functional coupling estimated from the pre-ictal periods of CHB-MIT scalp EEG and Xuanwu sEEG recordings, rescaled to model units.
The per-node estimator. Each candidate driver node receives sparse, low-amplitude probe pulses at random phases while in the oscillating state. For each probe we measure the phase shift over the next few cycles and fit a PRC with two or three Fourier harmonics. Because the fit is a Fourier series, its derivative, and so the predicted ARC, is available in closed form. The target phase is the minimum of the predicted ARC. The probe and quench phases are estimated causally with a delay-compensated phase extractor modeled on Shin et al., so latency errors are part of the evaluation rather than assumed away.
Driver selection. At each decision step, drivers are re-ranked by their outgoing coupling in the current sliding-window estimate of $W(t)$. The static-connectivity ablation instead uses a single time-averaged $W$.
The analysis questions. Three quantities organize the analysis. The first is how well the probe-derived phase matches the true optimal quench phase, found by brute force in simulation. The second is how far and how fast that optimal phase moves as $g$ drifts. The third is whether the width of the quench window shrinks as the node approaches the bifurcation, which would be the network analogue of Taylor's narrow window.
Baselines
Do nothing. No stimulation, so any seizure ends only through noise or internal dynamics. This gives the spontaneous termination rate and duration against which all controllers are measured.
Asynchronous periodic pulses. Pulses are delivered at a fixed rate regardless of phase, as a simplified stand-in for open-loop stimulation, and matched in total charge to our method. Mirkhani et al. argue that phase-random pulses partly cancel their own effects.
Continuous pinning feedback. This is the Liu et al. controller $u_i=-q_ix_i$ on the same driver nodes, re-derived for our nonlinear directed network and tuned by grid search over $q_i$ under the same budget used to tune our method. On its own fixed 8-node network, Liu et al. report that controlling 3 nodes quenched oscillations at all 8 nodes. For the delivered charge or energy needed to quench, there is no reported result found on Lacuna. We will therefore report both charge-to-quench and success rate, and count our method as winning if it matches the success rate at lower charge, or beats it once $W(t)$ varies.
Fixed-phase closed-loop stimulation, tuned once. This is the strongest practical competitor. Pulses target a single phase, as in Shin et al., but that phase is chosen by brute-force search at the start of each run, which gives it an advantage no real device has. It shares our phase extractor. The only reported number for this family is Shin et al.'s phase-tracking fidelity of above 95% correlation with an ideal Hilbert transform; no quench-efficacy number is found on Lacuna.
Direct amplitude search. A model-free Bayesian-optimization bandit searches over pulse phase, using measured amplitude reduction as its reward, with the same number of pulses our method spends on probes.
Ablations. Two ablations isolate the key ingredients. One replaces the derivative-based target phase with the PRC's own extremum. The other freezes driver selection at the time-averaged graph.
Why not just…?
Why not just tune one stimulation phase per patient and keep it? This is the real threat. If the optimal quench phase barely moves as inhibitory gain and coupling drift, a once-tuned fixed phase is simpler and just as good. Our second experiment measures this drift directly. We will report the range of $g$ over which the fixed phase stays inside the quench window. If that range covers the drift seen when $W(t)$ follows sEEG-derived trajectories, we conclude that adaptivity is unnecessary and say so.
Why not just use continuous feedback, as the pinning theorems assume? Clinical devices deliver pulses, not continuous currents, and Liu et al.'s guarantee rests on a static Laplacian eigenvalue that dynamic coupling violates. Whether this matters in practice is empirical. The time-varying-$W$ condition shows whether continuous pinning fails when coupling dips, and the charge comparison shows its cost when it succeeds.
Why not skip the derivative trick and search directly on amplitude? Amplitude reduction is exactly what we want, so a bandit on amplitude is the honest model-free rival. Our bet is that phase shifts are measurable from small, harmless probes, whereas amplitude effects near the separatrix are noisy and a bad probe can prolong the seizure. We compare sample efficiency and count iatrogenic prolongations for both methods. If the bandit matches us at equal probe budget, the derivative relation offers no practical advantage.
Proposed Experiments
Experiment 1: does the derivative relation hold for bistable nodes? In single Wilson–Cowan nodes swept across inhibitory gain $g$, we compare the ARC predicted from the PRC with the true ARC, using Pearson correlation. Mirkhani et al.'s $R=-0.84$ serves as a reference, though their setting differs. We also compare the predicted quench phase with the brute-force optimum. The decisive check is whether the predicted phase falls inside the true quench window across the bistable range. If it does not, we stop and report the failure.
Experiment 2: how much does the quench window drift? We measure how the location and width of the window move with $g$, including in the Taylor et al. thalamocortical setting, where the 0.15 s versus 0.20 s outcomes give a qualitative check.
Experiment 3: network control. We compare all baselines on the 8-node network of Liu et al. and on 15- to 18-node networks matched to the Xuanwu and CHB-MIT montages, under both static and data-derived $W(t)$. Metrics are quench success rate, charge-to-quench, time-to-quench, and the rate of pulses that prolonged oscillation. Each metric is reported under injected phase-estimation latencies of 0 to 20 ms. Seizure onset is signaled by an oracle, so detection is not confounded with control. A later run uses a detector front end; for reference, MB-dMGC-CWTFFNet reports an AUC of 0.935 and 0.059 false predictions per hour on CHB-MIT.
Experiment 4: physical testbed. We port the estimator and controller to the analog birhythmic circuit testbed of Birhythmic Analog Circuit Maze. The test is whether probe-derived timing survives real analog noise and latency.
What would count as success. The method succeeds if it beats the once-tuned fixed-phase baseline on success rate or charge under drifting $g$ and $W(t)$, and ties it when both are held fixed. Anything less means the adaptive machinery is not earning its complexity.
References
- Response of Neuronal Populations to Phase-Locked Stimulation: Model-Based Predictions and Validation
- Synchronization and Control in Bistable Oscillator Networks: Towards Epilepsy Regulation
- Abatement of epileptic spike-wave discharges through single pulse stimulation
- A 16-Channel Low-Power Neural Connectivity Extraction and Phase-Locked Deep Brain Stimulation SoC
- Predictive Coding of Dynamical Variables in Balanced Spiking Networks
- Efficient coding in biophysically realistic excitatory-inhibitory spiking networks
- Primate neocortex performs balanced sensory amplification
- Dynamic Multi-Graph Convolution-Based Channel-Weighted Transformer Feature Fusion Network for Epileptic Seizure Prediction
- Learning Dynamic Graph Representation of Brain Connectome with Spatio-Temporal Attention
- Brain Networks Should Be Learned, Not Constructed
- Birhythmic Analog Circuit Maze: A Nonlinear Neurostimulation Testbed