One EEG, Two Stories: The Hidden Lens in Brain-Flow Monitoring

University of Tokyo / EEG / Hypoperfusion / TAVI / Hilbert–Huang Transform / Multitaper

Make the heart race for a few seconds and blood pressure falls; less blood reaches the brain. University of Tokyo researchers passed one recording of that moment through two analytical methods. The images changed. So did the trajectory of the 95% spectral edge frequency. A monitor’s number, the work shows, is never the brain itself: between data and decision stands a method.

Ten seconds is brief on the wall clock of a catheter laboratory. Inside those ten seconds, however, the heart, arterial pressure and the electrical life of the brain can all change shape. A clinician sends stimuli through a wire in the ventricle, driving the heart faster than 180 beats a minute. The chambers contract so quickly that they eject little blood. Flow across the aortic valve quiets. In that controlled opening, a replacement valve can be expanded without being pushed out of position.

The maneuver is rapid ventricular pacing. It is used during transcatheter aortic valve implantation or replacement—TAVI or TAVR—to steady balloon dilation and valve deployment. No one lowers cerebral perfusion merely to conduct an experiment. The therapeutic step itself creates a short, clinically managed fall in pressure and brain blood supply, a “natural experiment” with an unusually distinct beginning and end.

Assistant Professor Ryo Wakabayashi, Professor Kanji Uchida and colleagues at the University of Tokyo Hospital’s Department of Anesthesiology and Pain Relief Center revisited frontal electroencephalograms recorded through that moment. They made two time-frequency pictures from the same voltage series. One used the multitaper method systematized by David J. Thomson in 1982. The other used the Hilbert–Huang transform, or HHT, which couples empirical mode decomposition with Hilbert spectral analysis and was introduced by Norden E. Huang and colleagues in 1998.

The university’s announcement on July 24, 2026, did not declare a universal winner. Its central finding was more fundamental: the same EEG looked different depending on the analysis, and the time course of the 95% spectral edge frequency, SEF95, differed as well. The colors and lines on a clinical screen are not a landscape directly photographed by an electrode. They are a landscape reconstructed through a window, a bandwidth, a smoothing rule and a mathematical model.

−30 to +60 secRegistered observation window around the start of rapid pacing

250 HzRegistered sampling rate for the frontal BIS EEG waveform

Every 3 secInterval for numerical data including blood pressure

95%Share of spectral power below the SEF95 boundary

90 peopleTarget enrollment in the registry—not a disclosed analysis count

2 methodsHHT and multitaper applied to the same EEG waveforms

What the public record establishes

The project was registered in the University Hospital Medical Information Network, UMIN, as an observational study of adults aged 20 or older who underwent TAVR under general anesthesia, had frontal EEG recorded with a BIS brain-monitoring system and experienced rapid ventricular pacing. Recordings with missing EEG data or artifacts were to be excluded. The registered primary outcome was the “specificity, rapidity, and interpretability” of HHT analysis in detecting cerebral ischemia. The study was self-funded, with a target sample of 90.

The analytical window ran from 30 seconds before pacing began until 60 seconds afterward. Numerical signals such as blood pressure were recorded every three seconds; EEG waveforms were captured at 250 samples per second. HHT was to calculate instantaneous energy in each frequency band. A multitaper time-frequency analysis would then process the same waveform, with spectral measurements including SEF95 compared between them.

The University of Tokyo ethics approval is dated October 2, 2023. The registry became public on December 24, 2024, and marks the study completed, with last follow-up on December 27, 2025. At the Japanese Society of Anesthesiologists’ 2025 annual meeting, a team comprising Yuka Tatematsu, Ryo Wakabayashi, Seiichi Azuma, Masaaki Asamoto and Kanji Uchida won the best-presentation award for work on HHT detection of cerebral ischemia. The work also appeared in English at the IARS/SOCCA annual meeting, and Wakabayashi has a 2025–26 research project to prototype a real-time hypoperfusion monitor based on HHT.

Known, reported and not yet disclosed

Registered: the operation, 90-second window, 250-Hz EEG, three-second numerical series, two methods, SEF95, target of 90 and observational design.
Reported by the university: the same EEG produced method-dependent time-frequency appearances and SEF95 trajectories.
Not disclosed: actual enrollment and analysis counts, exclusions, demographics, effect sizes, confidence intervals, sensitivity, specificity and clinical outcomes.
Not established: that one method is always superior, that the analysis prevented stroke, or that a real-time alarm improves outcomes.

The UMIN results section still read “unpublished” at its last update and did not list participant flow or outcomes. The university announcement did not link a peer-reviewed paper. It would therefore be inaccurate to turn the target of 90 into the number analyzed, or to state—without the study figures—that HHT was categorically faster or that multitaper missed ischemia. This account distinguishes the registered protocol and announced comparison from the deeper body of science explaining why two legitimate methods can disagree.

A voltage series is not yet a picture

An EEG is a time-ordered record of voltage differences, typically measured in millionths of a volt, between electrodes on the scalp. Much of the signal arises when postsynaptic currents in large populations of similarly oriented cortical pyramidal cells align in space and time. One trace is not the firing of one neuron. It is the distant, summed shadow of coordinated tissue.

An expert can read the raw line, but it is difficult to watch hours of waveform while simultaneously judging amplitude, frequency, asymmetry, suppression and artifact during surgery. Processed monitors therefore break the signal into short segments and estimate how much power resides at each frequency. Time becomes the horizontal axis, frequency the vertical axis and power a field of color—a spectrogram, often called a density spectral array.

That is where interpretation enters. How long is each segment? How much do adjacent windows overlap? Which frequencies are filtered? How are electrocautery, facial muscle activity, pacemaker interference and a loose electrode rejected? Where are the color limits? Different decisions make different images from the same raw data. Analysis is not a transparent afterthought that removes noise and reveals untouched truth. It is an extension of measurement, deciding what to preserve and what to average.

From Caton’s rabbits to Berger’s human scalp

In 1875, Liverpool physician Richard Caton reported electrical fluctuations recorded with a galvanometer from the exposed brains of rabbits and monkeys. He connected slow oscillations with sleep and wakefulness and related changes in electrical activity to sensory stimulation. The leap to non-invasive human recording came from Hans Berger, a psychiatrist in Jena, Germany.

Berger made his first human recording on July 6, 1924. He improved his instruments and scalp technique for years before publishing in 1929. A rhythm near ten cycles per second appeared when a person rested with eyes closed and diminished with eye opening or attention. Berger’s wave became the alpha rhythm. The fact that brain waves changed with consciousness, sleep, anesthesia, epilepsy and injury made continuous monitoring imaginable.

By the late 1930s, temporary carotid occlusion in animals had shown that electrical potentials disappeared when blood supply failed. EEG later became an adjunct for detecting ischemia during carotid endarterectomy and cardiac surgery. Yet anesthetic drugs themselves slow and reorganize the EEG. Temperature, carbon dioxide, age, pre-existing brain disease and muscle artifact also alter it. “Slower equals ischemia” is not a sufficiently safe dictionary for the operating room.

A separate 136-year history of cerebral flow

In 1890, Charles Roy and Charles Sherrington observed changes in the brain’s vascular volume with sensory stimulation, asphyxia and drugs. They proposed that active brain tissue helps regulate its own blood supply—an early form of what is now called neurovascular coupling. In the 1940s, Seymour Kety and Carl Schmidt devised a way to quantify whole-brain blood flow and oxygen metabolism in people by following inhaled nitrous oxide through arterial and internal-jugular blood.

Transcranial Doppler, near-infrared spectroscopy, PET and MRI now measure different aspects of velocity, oxygenation and regional perfusion at different spatial and temporal scales. EEG measures none of those directly. It measures the electrical consequence of neuronal populations receiving—or failing to receive—the oxygen and glucose needed to sustain their activity. Between arterial pressure and EEG lie cerebral autoregulation, collateral circulation, oxygen extraction and metabolic demand.

A 1975 study of 22 people with acute cerebral infarction found that EEG frequency in the opposite hemisphere often remained normal despite lower regional blood flow, with no correlation between the two measurements. That did not make EEG irrelevant to perfusion. It showed that local flow, recording location, timing and function do not reduce to a one-to-one conversion table. Other work on global hypoperfusion has found quantitative EEG changes beginning roughly 10 to 15 seconds after a cardiovascular event, with rapid but frequency-dependent recovery once circulation returns.

A blood-pressure cuff is not a cerebral-flow meter, and an EEG is not a flow meter. Link all three in time, however, and they can reveal how the brain buffers a circulatory shock—and when its electrical activity begins to yield.

Why a few seconds of TAVI make a good methodological test

Rapid pacing in TAVI commonly lasts about five to ten seconds at 180 to 220 beats a minute and may lower systolic arterial pressure below 50 mm Hg. Low stroke volume reduces the flow that would otherwise displace a balloon or prosthetic valve. Circulation often recovers quickly when pacing stops.

In a study of 173 TAVI patients monitored with near-infrared spectroscopy, cerebral oxygen saturation fell significantly from 64% before rapid pacing to 55% afterward. Smaller studies have observed the same direction of change. A temporary saturation fall is not synonymous with stroke. TAVI’s neurological risks also include calcific or tissue microemboli, air, bleeding, atrial fibrillation and sustained hemodynamic collapse.

For signal analysis, the benefit is timing. Align every patient at the instant pacing starts, and the seconds before, during and after low flow can be superimposed. Anesthetic concentration and other conditions may remain comparatively stable across that short 90-second slice. Because the event starts and stops abruptly, it is highly nonstationary—exactly the kind of signal that exposes how an analytical method distributes resolution between time and frequency.

Multitaper: several windows to steady the estimate

Fourier analysis represents a complex waveform as a sum of sinusoids at different frequencies. When a finite piece of data is cut out, the waveform ends abruptly at the boundaries. Power that belongs at one frequency can spread into neighboring frequencies, an artifact called spectral leakage. Applying a taper—a window that softens the segment’s edges—helps, but any one window leaves a tradeoff between bias, variance and resolution.

Thomson’s multitaper method applies several mutually orthogonal Slepian windows, or discrete prolate spheroidal sequences, to the same data. It calculates a spectrum through each window and combines the estimates. Averaging these near-independent views produces a smoother, lower-variance spectrum that resists leakage and random noise.

The cost is deliberate smoothing. A longer temporal window separates nearby frequencies better but blurs a brief event across time. A shorter window follows the event more sharply but cannot distinguish frequencies as finely. The number of tapers and the time-bandwidth product change the image too. Multitaper is not simply “slow.” It makes explicit choices about how much averaging is needed to obtain a stable spectral estimate from a short, noisy sample.

Hilbert–Huang: let the waveform choose the pieces

The first stage of HHT, empirical mode decomposition, does not begin with a fixed dictionary of sinusoids or wavelets. It connects the signal’s local peaks and troughs, forms upper and lower envelopes, subtracts their mean, and repeats the “sifting” process. The result is a sequence of intrinsic mode functions, or IMFs, generally ordered from faster to slower oscillation.

A Hilbert transform is then applied to each IMF. Its instantaneous amplitude and phase are calculated, and the rate of phase change yields instantaneous frequency. The resulting Hilbert spectrum places energy in time and frequency without first forcing the waveform into a predetermined set of components. This adaptive structure can sharply localize a transient event such as rapid pacing.

Data-driven does not mean assumption-free. Results depend on the stopping criterion for sifting, how boundaries are extended, how many modes are kept and how noise is treated. “Mode mixing” can put far-apart frequencies into one IMF or split one oscillation across several IMFs. “End effects” can create anomalous peaks near the boundaries of a finite record. HHT can increase the apparent magnification of a transient, but its lens has characteristic aberrations.

A 2023 study of propofol anesthesia in 30 patients illustrates both the appeal and the need for comparison. Instantaneous frequencies in HHT’s first two IMFs tracked movement from beta toward alpha activity during induction and reversed during emergence; IMF2 was associated with BIS. SEF95, by contrast, varied widely and did not change significantly during induction. Even when researchers ask about the broad construct “depth of anesthesia,” different summaries of one EEG do not necessarily move together.

SEF95, the single-number waterline

SEF95 is the frequency below which 95% of the total spectral power in a defined range lies. If power shifts toward slower activity, the line usually falls; if more power moves to higher frequencies, it rises. Displayed as a white trace across a spectrogram, it gives clinicians a quick view of trajectory, including possible left-right differences.

Its convenience comes from compression. Dozens or hundreds of spectral values collapse into one number in hertz. Very different spectra can share the same SEF95. It does not directly report total power, and it depends on the analyzed frequency range, filters, window, artifacts and spectral estimator. Facial electromyography can pull it upward. Anesthetics, hypothermia, age, sleep and hypoperfusion can all pull it downward.

The important Tokyo finding is that SEF95 trajectories derived from the same source EEG differed by method. SEF95 is not a unique substance in nature waiting for the electrode to collect it. It is a downstream quantity built from an estimated spectrum. If a clinical threshold depends on the line, the algorithm defining the denominator is part of the medical instrument.

Question	Multitaper	Hilbert–Huang transform
Basic idea	Estimate and combine Fourier spectra through several orthogonal tapers.	Extract data-adaptive IMFs, then calculate instantaneous frequency and energy.
Strength	Stable, low-variance spectral estimate with reduced leakage.	Sharp localization of nonstationary transitions without a fixed basis.
Main tradeoff	Window, bandwidth and taper count control time-frequency resolution and smoothing.	Sifting, mode mixing, end effects and noise sensitivity affect decomposition.
Clinical emphasis	“Where did the overall frequency distribution move?”	“When did the change occur and how did local oscillations deform?”
Tokyo conclusion	The same EEG produced different appearances and SEF95 time courses; public material does not establish universal superiority.

Why two valid calculations can disagree

First, a brief event cannot be measured with arbitrarily fine time and frequency resolution at once. Sharpen when something happened and its frequency content becomes broader; separate close frequencies and a longer observation interval is needed. Multitaper manages this tradeoff through the time window and bandwidth. HHT adapts to local waveform shape and frames the constraint differently, but it cannot manufacture information absent from a finite, noisy record.

Second, the methods answer different questions. Multitaper averaging is powerful when the goal is a stable estimate of band power hidden under random variability. HHT’s instantaneous frequency is attractive when an oscillation slides, divides or vanishes over seconds. A smoothed topographic map and a live map of water currents can describe the same valley without sharing every contour.

Third, preprocessing and edges matter. The registered Tokyo observation is 90 seconds long, but the duration of each computational segment and the baseline used for normalization can move or widen an apparent transition. HHT distributes the sharp beginning and end of pacing across modes. Multitaper distributes them across a window. SEF95 sits downstream and inherits those differences in its single line.

In the clinic, “look at both” may be the answer

The practical implication is not that every monitor should discard Fourier analysis for HHT. An acute-hypoperfusion alarm needs sensitivity to early change and specificity against electrocautery, facial muscle activity and other artifacts. A false negative can miss a threatened brain. A false positive can interrupt a procedure, prompt unnecessary vasopressors or trigger costly tests. A vivid display is not a clinical monitor until its output is tied to a useful action and a patient outcome.

A realistic approach is multimodal: inspect the raw waveform, spectrogram, SEF95, suppression ratio and asymmetry alongside blood pressure, the pacing marker, cerebral oxygen saturation, anesthetic concentration, temperature and carbon dioxide. HHT might flag the timing of a rapid transition while multitaper confirms a stable shift in background power. In that role, the methods are not rivals but audits of one another.

The next generation of alarms will need preregistered thresholds and external validation across hospitals—not merely the most attractive image. Researchers must test TAVI as well as carotid surgery, cardiopulmonary bypass, shock in intensive care, post-cardiac-arrest care and subarachnoid hemorrhage, then connect detection to stroke, delirium, cognition and survival.

What this study has not shown

Rapid-pacing hypotension is not identical to a focal thrombotic stroke. A few frontal electrodes may miss an occipital, deep or small unilateral lesion. EEG obtained through a BIS device may also have device-specific preprocessing and is not necessarily identical to a raw research amplifier signal.

TAVI patients are often older and have severe aortic stenosis. Anesthetic regimen, valve type, number and duration of pacing episodes, baseline pressure, ventricular function, carotid disease and hemoglobin could all modify the response. If this is a single-center observational dataset, younger people and other causes of hypoperfusion require separate validation.

Most importantly, the university summary establishes method-dependent differences but does not quantify which differences were clinically important. A peer-reviewed report, participant flow, analysis code, parameter settings and the handling of artifacts are needed before another center can reproduce the images or compare performance.

What the next study should provide

A participant flow diagram with actual enrollment, exclusions and analysis counts.
Patient-level links among pacing duration, minimum mean pressure, cerebral oxygenation and EEG response.
Full disclosure of window length, bandwidth, taper count, EMD stopping rules, boundary handling and filters.
Outcome-blinded readers and preregistered detection latency, sensitivity and specificity.
Simultaneous multichannel EEG, near-infrared spectroscopy and transcranial Doppler.
Shared code and representative de-identified waveforms, followed by external replication.

In the AI era, display the method’s name

Future EEG monitors may add machine learning and display “68% risk of hypoperfusion.” Yet AI also chooses—or inherits—the spectrum and features it sees. A model trained on HHT maps and a model trained on multitaper maps can return different probabilities for the same patient. If training data, preprocessing, missing-data rules and device generation are hidden, the number may appear precise while its meaning remains unstable.

The Tokyo comparison therefore reaches beyond two signal-processing techniques. It supports a larger rule for medical AI: keep raw data, transformation, summary metric and decision model conceptually separate, and record uncertainty at every layer. A clinical screen should expose the method and signal quality, and perhaps show whether an alternative analysis agrees. Disagreement need not be concealed as failure. It can be information telling the clinician to pause.

From “which is true?” to “what did each preserve?”

Caton’s galvanometer showed that the brain has an electrical life. Berger showed that human state could be read from the scalp. Roy and Sherrington, then Kety and Schmidt, made the relation between activity and blood supply a measurable problem. Fourier, Thomson and Huang provided different mathematical routes from a trembling line to a map of frequency through time.

After 150 years, no instrument has become a transparent window onto the brain. In fact, the more polished the display, the more important it is to understand the glass. The University of Tokyo result does not reveal an unfortunate defect—that analysis “interferes” with a pure answer. It reveals that analysis is part of the measurement system that produces the answer.

When flow falls during surgery, clinicians do not truly need to know which beautiful spectrogram is metaphysically correct. They need to know what changed in neuronal activity, when, by how much and whether restoring circulation reversed it. Agreement between methods may identify a robust signal. Disagreement may expose the event’s time scale or the uncertainty in the estimate. When one EEG tells two stories, science should not erase one. It should measure why they differ, then turn that difference into safer care.

Sources and references

This article is based primarily on the University of Tokyo’s July 24, 2026 announcement and the UMIN registry, supplemented by primary and peer-reviewed work on rapid pacing in TAVI, cerebral oxygenation, EEG history, cerebral-flow measurement, multitaper analysis, HHT and SEF95. It does not infer an analysis count or effect size absent from public records.