Some of the most important physiological signals in medicine are also among the hardest to measure. Intracranial pressure (ICP) and cerebral blood flow (CBF) are central to managing traumatic brain injury, stroke, and other neurocritical conditions. Yet the standard way to know a patient's ICP is to drill through the skull and place a probe. It works, but it is invasive, it carries risk, and it is reserved for the sickest patients in the most controlled settings. For most of the people who could benefit from knowing their brain's state, we simply don't measure it.
The question that drove a large part of my work at Liminal and Hyperfine was deceptively simple: can you estimate what's happening inside the skull without going inside the skull? The approach we developed — what we came to call Acousto-Encephalography (AEG) — is one answer to that question.
The signal is never the thing you want
The brain is one of the noisiest, most confounded systems you can try to sense from the outside. The skull attenuates and distorts. Tissue is heterogeneous. The quantities you actually care about — pressure, perfusion — are not directly observable. What you can measure are correlates: how acoustic waves propagate and reflect, how blood pulses through vessels, how multiple modalities co-vary over time.
This is the central difficulty, and it generalizes far beyond the brain. You rarely get to measure the latent state directly. You measure something downstream of it, entangled with everything else, and your job is to invert that mess into an estimate you can trust.
AEG was never really about a single transducer or a single algorithm. It was about treating the skull-and-brain as a system and asking what combination of physics and learning could recover a hidden state from indirect evidence.
So AEG was deliberately multimodal. Ultrasound carries information about mechanics and flow; combined with other sensing and with the temporal structure of the cardiac and respiratory cycles, it becomes possible to constrain estimates that any single modality leaves ambiguous. The information isn't in any one channel — it's in how the channels constrain each other.
It helps to make this concrete. A single ultrasound pulse, on its own, is wildly ambiguous: a hundred different combinations of skull thickness, tissue properties, and underlying pressure could have produced the echo you just recorded. But those combinations are not all equally consistent with the cardiac pulse you measured a moment earlier, or with the way the waveform shifts as the patient breathes, or with the slow drift you've been tracking over the last ten minutes. Each additional constraint carves away part of the space of explanations. No single measurement identifies the state; the intersection of many of them does.
A short detour on what ICP and CBF actually tell you
It's worth being precise about why these two numbers matter so much. Intracranial pressure is, in effect, the pressure of the closed box the brain lives in. The skull does not expand, so when something inside it swells — bleeding, edema, a mass — pressure rises, and rising pressure squeezes the very blood supply the brain depends on. Cerebral blood flow is the other side of that coin: how much oxygenated blood is actually reaching tissue. Clinicians care about both because the dangerous failures are dynamic. A patient can look stable while their compensatory reserve quietly runs out, and then deteriorate fast. The decisions that follow — when to intervene, how aggressively, whether to escalate — hinge on knowing not just a single value but a trend.
This is exactly why a one-time, invasive measurement is such an awkward fit for the problem. The thing you most want is continuous visibility into a quantity that is moving, and the tool you have is a risky snapshot you can only justify for the sickest patients. AEG was an attempt to close that gap from the other direction: trade a little precision for the ability to watch the trajectory, non-invasively, on far more people.
Physics and learning, together
There is a long-running tension in sensing between physics-based models and data-driven models. Physics-based models encode what we know — wave propagation, tissue interaction, the mechanics of perfusion. They are interpretable and they generalize, but they break against the messiness of real biology and real patients. Data-driven models absorb that messiness, but they can be brittle, opaque, and dangerously confident outside their training distribution.
The most useful systems I've built refuse to choose. Physics gives you the scaffolding: the forward model of how a hidden state produces the measurements you see. Learning fills in what the physics can't capture: patient-specific variation, unmodeled confounds, the residual structure in real data. AEG sat firmly in that hybrid space — first-principles understanding of how waves move through the skull, coupled with models that adapt to the variability you can't write down in an equation.
This matters for a clinical context for a specific reason: interpretability and safety are not optional. A black box that outputs an ICP number is not deployable if no one can reason about when it will be wrong. Anchoring the system in physics gives you something to check the learning against, and a story for why an estimate should be believed.
The practical payoff of the hybrid approach is that the two halves cover each other's failure modes. When the data-driven component drifts into a regime it hasn't seen, the physics acts as a sanity rail: an estimate that violates conservation or implies a physically impossible wave path can be flagged or down-weighted. When the physics model is too idealized to match a particular patient's anatomy, the learned residual absorbs the discrepancy instead of letting it leak into the final number. Neither half is trustworthy alone. Together they degrade gracefully instead of catastrophically — which, in a neuro-ICU, is the only kind of degradation you can live with.
Designing under real clinical constraints
Building a sensor that works on a bench is a different problem from building one that works in a neuro-ICU. The constraints reshape everything:
- You don't control the environment. Patients move, equipment varies, the acoustic coupling is never ideal.
- Variability is the norm. Skull thickness, anatomy, and physiology differ enormously between people; a system tuned to one population fails on another.
- Failure modes have consequences. A wrong estimate isn't a bad demo — it's a clinical decision made on bad information. The system has to know when it doesn't know.
These constraints pushed the work to be end-to-end. It wasn't enough to design a clever transducer. We had to think about signal acquisition robust to poor coupling, processing pipelines that held up under noise and patient variability, models that mapped measurements to clinically meaningful representations, and the embedded and cloud infrastructure to run inference in something close to real time. Every layer had to be co-designed, because a weakness in any one of them showed up as an unreliable number at the end.
The coupling problem deserves special mention, because it is the kind of thing that looks trivial in a paper and dominates your time in practice. Ultrasound only tells you something if the sound actually gets into the head. A few millimeters of air gap, a patient who shifts, a careless application — and the signal you're so carefully inverting is mostly an artifact of the interface, not the brain. So a large part of "the algorithm" is really about detecting when the measurement is untrustworthy in the first place: recognizing poor coupling, motion, and out-of-distribution conditions, and refusing to emit a confident number when the input doesn't earn one. A sensor that stays silent when it's uncertain is far more useful in a clinical setting than one that's confidently wrong a few times a shift.
Calibration, and the long tail of being human
One of the quieter lessons of AEG is how much of the difficulty is concentrated in the tails of the population rather than the middle. It is not especially hard to build a system that works on a typical adult skull under good conditions. It is very hard to build one that also holds up on the unusually thick skull, the unusual anatomy, the patient whose physiology sits at the edge of what your models have seen. And those tails are not rare curiosities you can wave away — in a clinical population they are a meaningful fraction of the people you most need to help.
This reframes what "accuracy" even means. A single headline number — average error across some test set — hides exactly the behavior that matters, because the average is dominated by the easy middle of the distribution. The questions that actually determine whether a system is deployable are about the tails and the failure modes: How does it behave on the patients it has seen the least? Does its confidence collapse appropriately when it's out of its depth, or does it stay falsely high? When it's wrong, is it wrong by a little or by a lot, and does it know? A system that is excellent on average and silently catastrophic on the hard 5% is not a system you can put in front of a clinician.
So calibration — the system's estimate of its own uncertainty — ends up being as important as the point estimate itself, maybe more so. A well-calibrated AEG estimate that says "I'm confident this ICP is elevated" and a well-calibrated one that says "I genuinely don't know, get a direct measurement" are both useful, because both let the clinician act correctly. An overconfident wrong estimate is the only truly dangerous output, and most of the engineering discipline goes into making sure the system earns its confidence before it expresses it.
Why this points toward closed loops
AEG is a sensing technology, but the reason it excites me is what it enables downstream. Once you can continuously and non-invasively estimate a brain state, you are no longer taking a one-time snapshot — you are tracking a trajectory. And once you can track a trajectory, you can start to close a loop: detect deterioration earlier, titrate intervention against measured response, and eventually couple sensing to modulation.
That is the throughline connecting this work to everything else I do. Whether the signal is intracranial pressure, neural activity, or a person's voice, the pattern is the same: weak, indirect, confounded measurements on one side; a hidden state you actually care about on the other; and a system in the middle that has to bridge them well enough to act. AEG was one of the clearest lessons I've had in how hard — and how worthwhile — that bridge is to build.
There's a broader point hiding in here about where medical progress comes from. We tend to imagine it arriving as new therapies — a drug, a device, a procedure. But a great deal of it comes, less visibly, from making the invisible measurable. Every time we turn a quantity that used to require a risky, expensive, specialist procedure into something you can estimate cheaply and continuously, we change who gets monitored, how early problems are caught, and what kinds of treatment loops become possible at all. AEG is one attempt at that move for the brain. The harder and more general lesson — that the bottleneck is rarely the model and almost always the system that turns a confounded signal into a decision you can stand behind — is the one I've carried into everything since.