Why Timing May Matter More Than Power

In a 2018 study published in Nature Communications, Youssef Ezzyat and colleagues worked with twenty-five patients who already had electrodes implanted in the brain for epilepsy monitoring. While the patients memorized lists of words, the researchers recorded activity from temporal and medial temporal regions. A classifier trained on earlier sessions produced a running estimate of whether the current network state was likely to support later recall. When that estimate fell below a preset threshold — when a memory looked likely to fail — the system delivered a brief pulse of electrical stimulation to lateral temporal cortex.

Stimulation applied on a fixed schedule, without regard to brain state, had produced inconsistent results in earlier work. Here, triggering only at the unfavorable moments modestly improved recall, raising the odds of remembering a word by roughly eighteen percent relative to no stimulation, while control stimulation at other sites did not help. That is the shape of the claim this article examines. No one has shown that a memory can be written into a brain. What the experiment shows is that an intervention can gain leverage by arriving at the right moment.

It would be a mistake to read the result as proof that memory has been decoded or that a prosthesis is imminent. The effect was small. The classifier was right only somewhat more often than chance, with an area under the curve near 0.61. The participants were epilepsy patients with intact memory, not people with amnesia, and the study measured recall of word lists rather than anything resembling daily life. What the experiment demonstrates is narrower and more useful: timing changed the outcome of a fixed dose of current, and the explanation for that belongs to the biology rather than to the electronics.

What “the right moment” assumes about nervous tissue

A neuron does not respond to an input with a fixed reflex. Its willingness to fire depends on what it was doing immediately beforehand. After an action potential, the cell passes through a refractory period during which it is briefly resistant to excitation. The membrane potential drifts between states that favor and oppose firing. Inputs from thousands of other cells arrive with their own rhythms. The result is that excitability fluctuates, and an identical stimulus applied at two different phases of an ongoing rhythm can produce two different responses, or none at all.

This principle is not confined to one structure. In motor cortex, the phase of the sensorimotor mu rhythm and of the beta rhythm shapes how readily a magnetic pulse over the scalp evokes a twitch in a hand muscle. In the hippocampus, the phase of the theta rhythm helps determine whether a synapse strengthens or weakens. In thalamocortical circuits, whether a click is woven into a spindle or experienced as a startle depends on when it lands. Physiologists call this state dependence, and it is the reason a fixed dose of stimulation has a variable effect.

Once state dependence is taken seriously, the engineering picture changes. Suppose a device must move a network out of an unwanted state and toward a better one. One strategy is to apply energy continuously and forcefully enough that the unwanted state cannot persist. That is roughly how conventional deep brain stimulation has operated: steady, high-frequency pulses delivered all day, whether the circuit needs them or not. A second strategy is to observe the circuit, wait for the moment when it is most responsive or most vulnerable, and deliver a small input then. The second strategy is harder to build because it requires real-time measurement and a defensible model of what the measurement means. Its promise is that it might achieve comparable effects with less energy, fewer side effects, and less interference with the brain’s own processing.

The clearest human evidence concerns memory during sleep

The most direct support for timing as a therapeutic lever comes from sleep research, where the relevant rhythm is slow, large, and comparatively easy to track in real time.

In 2013, Hong-Viet Ngo and colleagues reported a closed-loop auditory experiment with eleven participants. Slow oscillations during deep sleep have a natural cycle of roughly one second, and their “up” phases are when cortical neurons are most excitable. The team detected the rhythm as it unfolded and played brief pulses of pink noise in phase with the up states. In-phase stimulation enhanced the slow oscillation, boosted the fast spindles coupled to it, and improved overnight retention of word pairs — roughly twenty-two words recalled the next morning, against thirteen in the control condition. Stimulation delivered out of phase did nothing. The timing carried the effect; the amount of sound did not.

Three years later, Caroline Lustenberger and colleagues applied the same logic to a faster rhythm. Sleep spindles are transient bursts of eleven to sixteen hertz activity, and their contribution to memory had been inferred mostly from correlations: people who spindled more tended to remember more. The team built a system that detected spindles in real time and applied brief epochs of transcranial alternating current stimulation at spindle frequency, but only when a spindle was already underway. Across sixteen participants in a randomized crossover design, stimulation increased spindle activity and improved overnight consolidation of a motor sequence task. Word-pair memory, a declarative measure, did not improve. Again, the benefit depended on matching the input to an event the brain had already initiated.

The counterexample matters as much as the positive results. In 2019, Nicholas Henin and colleagues tried to reproduce the auditory intervention and found no behavioral benefit. Their study enhanced slow oscillations and spindle activity — robustly, in two separate cohorts — but memory did not improve, and Bayesian analysis favored the absence of an effect. The lesson is uncomfortable. A closed-loop system can demonstrably change the rhythm it targets without changing the behavior anyone cares about. Entrainment is not benefit. A measurable oscillation may be a correlate of a memory process rather than its cause, and moving the correlate is not the same as moving the process.

Parkinson’s disease is the most advanced test case

If timing-based stimulation has a proving ground, it is Parkinson’s disease, because the pathological rhythm is prominent, the electrodes are already implanted for therapy, and the outcome can be rated by a clinician who does not know which condition the patient received.

In 2013, Simon Little and colleagues tested adaptive deep brain stimulation in eight patients. Rather than stimulating continuously, the device monitored beta-band activity in the subthalamic nucleus and switched stimulation on when the amplitude crossed a threshold, typically within thirty to forty milliseconds. Blinded motor ratings improved substantially — around half again better than conventional stimulation — while the total time stimulation was actually delivered fell by more than half. The study was small, the trials lasted minutes rather than months, and it was a proof of concept rather than a finished therapy. It nonetheless showed the shape of the claim: less stimulation, delivered selectively, could outperform more.

Later work clarified the mechanism. Beta activity in Parkinson’s disease does not sit at a permanently elevated level; it arrives in bursts of varying duration. Gerd Tinkhauser and colleagues demonstrated in 2017 that the amplitude of beta activity grows with burst duration, as if more and more neurons were being recruited into a synchronized pattern, and that long bursts tracked worse motor impairment. Adaptive stimulation truncated the long bursts and shifted the distribution toward short ones. Conventional continuous stimulation suppressed beta activity globally without altering the burst structure. Two interventions that both “lower beta” do different things to the circuit.

The story has not resolved cleanly in favor of timing. In 2024, a blinded randomized crossover trial led by Carina Oehrn tested chronic adaptive stimulation in four patients and found that time spent with the most bothersome symptoms fell by roughly half compared with optimized conventional stimulation. But the biomarker that worked best was not beta; it was stimulation-entrained gamma activity measured in the subthalamic nucleus or motor cortex. Commentators also observed that the adaptive algorithm delivered more total stimulation than the conventional setting, including during sleep, so the trial did not simply demonstrate that less is more. With four patients, it could not.

Magnetic pulses and the phase controversy

Non-invasive methods test the same principle with a cleaner instrument. A magnetic pulse over the motor cortex evokes a small twitch, and the size of that twitch indexes cortical excitability at that instant. If excitability really depends on phase, the twitch should wax and wane as the pulse timing moves around the ongoing sensorimotor rhythm.

Several studies report exactly that. In 2022, Wischnewski and colleagues described a double dissociation: pulses timed near the trough and rising phase of the mu rhythm produced larger responses, while the beta rhythm pushed in the opposite direction. Earlier work by Til Ole Bergmann and colleagues found that the plasticity induced by repeated pulses depended on the rhythm’s state at the moment of stimulation, and other studies place peak excitability tens of degrees after the negative peak of the cycle.

The result is fragile, however. A 2019 study by Madsen and colleagues found no reliable phase dependence of corticomotor excitability. A 2022 replication attempt by Ozdemir and colleagues reported a phase effect only at high oscillation amplitude, and it did not reproduce across separate visits. The honest summary is that phase-dependent excitability is real in some preparations and elusive in others, and that the effect is small enough for measurement noise, individual anatomy, and analysis choices to swamp it. Anyone promising that a headset can “tune your brain waves” is skipping the part where the effect has to survive replication.

Why amplitude remains the default

Given all this, why do most devices still push a constant current?

Continuous stimulation is easy to reason about. It can be tested in a conventional randomized trial with a fixed dose. It does not depend on a classifier that might mislabel brain states, or on signal processing that inserts a delay between detection and intervention. Its regulatory pathway is familiar, and a fixed amplitude can be written on a label. Complexity is a cost, and it must be paid for with a benefit large enough to justify it.

Closed-loop timing introduces failure modes of its own. The measurement can be noisy. The model that maps a signal to a state can be wrong in ways that remain invisible until they matter. The detected state can drift over months as a disease progresses or a patient ages. There is a lag between sensing and stimulating, and if the target rhythm moves faster than the loop, the pulse arrives late. A loop that is tuned to deliver more stimulation whenever a biomarker drifts upward can quietly escalate the total dose. The Oehrn trial is a reminder that adaptive algorithms optimize whatever they are pointed at, and what they are pointed at is a proxy.

There is also a conceptual trap in the word “power.” Clinicians think in terms of amplitude, pulse width, and frequency, all of which can be specified and compared. Timing has no equivalent unit. A claim that one device is better because it is “smarter” is harder to audit than a claim about voltage, because the benefit depends on the accuracy of the state estimate, the quality of the biomarker, and the match between the algorithm and the individual. A field that cannot yet agree on which biomarker to use is a field in which “clever timing” can mean many different things.

What would make a timing claim credible

A claim that timing beats power should be judged on more than a compelling demonstration. It should be repeated in an independent sample. It should be compared against a condition that matches total stimulation dose, so that any benefit can be attributed to timing rather than to more current. It should improve something a patient can feel — walking, speech, sleep quality, the recall of a name — rather than only a spectral feature on a research amplifier. It should report adverse events candidly. And it should show that the benefit survives outside the laboratory, because a rhythm that can be steered for twenty supervised minutes is not the same as a device that improves a life.

The skepticism worth holding is not about the physics. Phase-dependent excitability is well established in principle. The open question is whether the effect is large and stable enough, in any particular application, to justify the added machinery. For sleep and memory, the evidence is genuinely mixed. For Parkinson’s disease, adaptive stimulation is promising and still unproven at scale. For consumer devices that claim to optimize cognition by phase-locking to an EEG rhythm at home, the demonstrated science runs out well before the marketing begins.

Demonstration, engineering, and speculation

It helps to keep three categories separate, because public discussion routinely blends them.

Demonstrated science, as of this writing, includes the following. Neuronal responses depend on the recent history and phase of ongoing activity. Closed-loop stimulation can enhance specific oscillations in humans. In some experiments that enhancement accompanies improved memory or motor performance, and in others it does not. Adaptive deep brain stimulation can reduce motor symptoms in small groups of Parkinson’s patients while using less energy, though the best biomarker and algorithm remain unsettled. Phase-dependent cortical excitability exists but is difficult to reproduce reliably.

Plausible engineering includes the following. Devices that estimate brain state in real time and intervene selectively could, within a decade, treat certain movement and memory disorders with fewer side effects than fixed stimulation. Wearable systems might detect states such as drowsiness or overload and deliver gentle interventions. Each of these depends on solving measurement, latency, and generalization problems that are currently unsolved at scale.

Speculation includes the vision of a general-purpose dial for mental performance: a headset that reads your rhythm and nudges it toward whatever the task requires, reliably and to order. Nothing in the current literature rules that out, and nothing in it demonstrates that either. The gap between a promising laboratory effect and a dependable product is where most neurotechnology claims quietly go wrong.

Where the leverage actually is

Timing matters because brains are not passive conductors of input. They are active, oscillating systems with their own internal schedules, and an input that ignores those schedules wastes most of its energy and pays for the remainder in side effects. The bet behind closed-loop stimulation is that listening first and acting second is worth the added difficulty. The evidence assembled so far — from memory consolidation in sleep, from beta bursts in Parkinson’s disease, from the contested phase dependence of cortical excitability — supports that bet as an engineering direction while leaving its size and reliability genuinely open.

There is a further reason to take the question seriously that has nothing to do with efficacy. A device that can identify and target a specific state is more selective than one that bathes a region in current, and selectivity is what makes the ethical questions about neurotechnology sharper rather than softer. The more precisely a system can read a person’s internal state, the more precisely it can act on that state, for better or worse. Whether the added leverage is used to widen a patient’s range of voluntary action or to narrow it for someone else’s convenience depends on decisions that no amount of engineering will make for us.