NeuroEdge Nexus

Part I of this series ended with a specific claim: that what clinical neuroscience is missing is not innovation, but integration. We followed a single discovery Llinás’s finding that neurons are self-sustaining electrical generators whose rhythms synchronise across distant regions of the brain — all the way to what it means for epilepsy surgery in 2026. The argument was that disease doesn’t live in a single anatomical place. It lives in the relationship between places, in the conversation between groups of neurons whose rhythms have fallen into an abnormal pattern. That shift has consequences well beyond epilepsy. This edition is about one of the most important of them.

There are two numbers worth holding at the start of this argument.

The Phase II and III failure rate for drugs targeting the central nervous system is approximately 85%. That figure has not moved substantially in decades. (BioSpace, 2024) On average, it takes seventeen years for research evidence to reach clinical practice, and only about 14% of it gets there. (Balas & Boren, 2000)

These are not numbers about bad science. The science is often very good. They are numbers about a system that was never fully designed to move what the laboratory understands toward the people who need it. And in neurology specifically, the point where that system breaks most consistently is not at the molecular end (where the tools have become extraordinary) but somewhere in the middle: in the step between what a genetic or molecular finding shows, and what it actually produces in the living network of a patient’s brain.

That gap is what this edition is about.

I want to explain where this argument comes from, because the timing matters — and because it is not biography.

During my residency in clinical neurophysiology at Ramón y Cajal University Hospital, I spent my late afternoons as an observer at the Cajal Institute . These are two separate institutions that share a name, and that name belongs to Santiago Ramón y Cajal — Nobel Prize in Physiology or Medicine 1906, the Doctor who described the neuron as the fundamental unit of the nervous system, whose illustrations remain scientifically accurate more than a century later.

The Cajal Institute holds his legacy: his original drawings, his instruments, the physical record of what one mind understood about the nervous system. The hospital that carries his name holds the patients.

In those months I moved between them daily. The researchers at the Cajal Institute were working at extraordinary depth, specific questions, rigorous methods, science that takes years to build. The patients I was learning to evaluate at the hospital had exactly the same pathologies those researchers were studying. Most of what I observed in those late afternoons would never reach those patients. Not because the science was wrong. Because the path connecting the two didn’t exist in any coherent form.

I mention this not as personal context but because it is a precise, verifiable illustration of the structural argument this article makes: two institutions carrying the same name, studying the same diseases, operating in separate worlds. Not a metaphor. A fact. And one that points directly at the nature of the problem — not a shortage of scientific depth, but an absence of the architecture that connects that depth to the patient.

That absence is why NeuroEdge Nexus exists. And it is the tension this article is about.

Before my residency in Madrid, during medical school in Bogotá, Dr. Fabio Andrés Rodríguez Cely taught me biochemistry, molecular biology, and genetics — not as a sequence of isolated chapters but as a single continuous thing. The molecule connecting to its function, the function to its physiological consequence, the consequence to what you eventually observe in a patient. He was, simultaneously, Medical Director at major pharmaceutical companies — a scientist who understood how industry worked from the inside, and who never let the molecule exist without its downstream meaning.

I raise this here — a teacher from more than twenty years ago — because what he was describing then is precisely what the translational pipeline of 2026 still does not do systematically. The integration of molecular biology, protein function, physiology, and clinical consequence as a single continuous chain of evidence, not as parallel tracks that occasionally intersect. The problem he was teaching against is the same problem the 85% describes. The timing is not coincidence. It is evidence that this is not a new failure.

What Llinás demonstrated — first at the level of a single neuron in 1988, then across whole-brain networks in 1999 — is that neurological disease is not primarily a molecular event. It is a network event. The same disrupted conversation between the thalamus and the cortex appeared across chronic pain, tinnitus, Parkinson’s disease, and depression — conditions that share almost no molecular features, but share the same functional network pathology. (Llinás et al., PNAS, 1999)

This has a direct implication for how we evaluate whether a therapeutic intervention works. If the disease lives at the level of the network conversation, then evidence that a drug changed a protein or a pathway is not the same as evidence that it changed the conversation. The molecular finding and the functional network finding answer different questions. One does not substitute for the other.

And yet the dominant translational framework is still built almost entirely around the molecular side of that equation.

Genetics tells you what could change. Proteomics tells you what is changing at the molecular level. Organoids model what might happen in human neurons under controlled conditions. These are meaningful advances — multi-omics integration, patient-derived iPSC models, AI-assisted target prioritization bring human biology earlier into the process, and in my opinion the direction is right.

But none of these tell you what is actually happening in the functional network of the living patient.

If we can analyse what is occurring at the genetic and molecular level in a neuron, but we are not simultaneously measuring the neurophysiological response of that neuron and its network, then we do not actually know whether the change we produced at the molecular level is the one generating the consequence we are aiming for. That chain of evidence has a gap exactly in the middle — between the molecule and the network function — and it is in that gap that most of the 85% disappears.

In clinical practice, that gap has a form that anyone who has worked in neurophysiology long enough will recognise.

I have evaluated patients whose molecular profile — amyloid burden, inflammatory markers, biomarker ratios — indicated advanced pathology, while their functional neurophysiology was more intact than those numbers would have suggested. And the reverse: patients with modest biomarker profiles showing severe network disorganisation on EEG, markedly abnormal evoked potentials across multiple modalities, functional evidence of breakdown that the molecular picture alone would not have predicted.

The molecule and the network were not telling the same story. In my experience, they often don’t. And the clinical decision — the real one, the one that matters to the patient — cannot be made from one without the other.

This is not a theoretical observation. It is what Rodríguez Cely’s integrated way of seeing predicts, and what Llinás’s network model of disease explains: that the molecular finding and its functional consequence are related but not identical, and that assuming one from the other is where the chain of evidence breaks.

The tools for measuring the network side of this equation already exist. This is not an argument for developing something new. It is an argument for where in the pipeline tools that already exist should be applied — and why their absence at early development stages is not a technical limitation but a structural choice with measurable consequences.

An EEG does not tell you which protein changed. It tells you whether the network changed — whether an intervention reached the level where, as Llinás showed, the disease actually lives. Designed into a trial from the beginning rather than added at the endpoint, it is not a monitoring instrument. It is a proof-of-mechanism instrument.

Evoked potentials — somatosensory, visual, auditory, motor — tell you whether a specific pathway is conducting and integrating. Not a molecule. A circuit, measurable and reproducible, regulatory-grade when properly designed into a protocol.

Single-unit neuronal recording goes further still: the firing pattern of an individual neuron within its local network, its temporal relationship to neighbouring cells. This is the scale at which a molecular change either becomes a functional change or disappears — where the protein story either translates into network behaviour, or doesn’t, and you need to know which before committing to a Phase III.

And repetitive TMS, beyond its therapeutic applications, offers something that remains underused as a research instrument: a non-invasive window into cortical excitability and plasticity in the living human brain. A 2025 paper in Neuron showed that high-frequency rTMS applied to primary motor cortex reverses the excitation-inhibition imbalance in the S1-M1 microcircuit, restoring glutamatergic neuronal activity — not by targeting a molecule, but by operating directly at the network level. (Wang et al., Neuron, 2025) That result matters here not primarily as a therapeutic finding, but as a demonstration that network-level change is measurable, non-invasively, in humans — and that molecular and functional evidence can be generated simultaneously, not sequentially. That simultaneity is precisely what the current pipeline is missing.

I have worked with these instruments across more than thirteen years of clinical practice. What they generate is a layer of evidence that no omics platform produces, and that regulators at both the FDA and EMA are increasingly asking for: real-world evidence that an intervention changed something meaningful in the patient’s nervous system, not only in a biomarker panel.

A 2024 study in Neurology found that 46% of Phase III trials for neurological diseases were launched without a preceding positive Phase II result. Programs that advanced without solid Phase II foundations showed a 31% positive outcome rate. Those built on them reached 57%. (Moyer et al., Neurology, 2024)

The difference between 31% and 57% is not explained by molecular science. It is explained by the quality of the translational evidence built before the Phase III commitment was made — and the most consistent absence in that evidence is functional network validation at the stages where it would most change decisions.

Both the FDA and EMA are asking for this with increasing specificity. That is not new regulatory preference. It reflects what the science has been pointing to since Llinás published in 1999, and what Rodríguez Cely was teaching long before that: that the molecule and its functional consequence are not the same thing, and that a pipeline which measures one without the other is not measuring what it thinks it is measuring.

The path from a discovery to a patient is long. It passes through scientific validation, regulatory review, institutional adoption, clinical implementation — each layer with its own logic and timeline. Most evidence does not survive all of them. Seventeen years. Fourteen percent.

That is not a failure of science. The science exists. The effort is genuine. What it describes is a gap in the evidence architecture — specifically, in the step between molecular discovery and functional network validation: between what happens in a gene or a protein, and what that change produces in the network of a living patient’s brain.

The wrong question is whether we have enough molecular evidence. We often do.

The right question is whether the evidence connects the molecule to the network to the patient — whether the change produced at the molecular level is actually the one generating the consequence we are aiming for, measured in a way that holds under regulatory scrutiny and translates into something a clinician can act on.

Cajal described the neuron. The Cajal Institute and the Ramón y Cajal Hospital carry his name and, more than a century later, still largely operate in separate worlds. The science in one building and the patient in the other. That separation is not inevitable. But closing it requires something the pipeline has not yet built systematically: the simultaneous measurement of what changes at the molecular level and what that change produces in the network.

That is the missing layer.

Part III will examine what building it would actually require — and why the barrier is not scientific but structural, a governance one

Part I examined thalamocortical dysrhythmia and what the network model of disease means for epilepsy surgery in 2026. Part III will examine what a regulatory-grade functional biomarker framework for neurological drug development would need to include — and why building it is less a question of scientific capability than of how the system is designed.

🔗 neuroedgenexus.com

References

1. Kola I, Landis J. Can the pharmaceutical industry reduce attrition rates? Nat Rev Drug Discov. 2004;3(8):711–6. DOI: 10.1038/nrd1470

2. Balas EA, Boren SA. Managing clinical knowledge for health care improvement. Yearb Med Inform. 2000;9(1):65–70. DOI: 10.1055/s-0038-1637943

3. Llinás RR. The intrinsic electrophysiological properties of mammalian neurons: insights into central nervous system function. Science. 1988;242(4886):1654–64. DOI: 10.1126/science.3059497

4. Llinás RR, Ribary U, Jeanmonod D, Kronberg E, Mitra PP. Thalamocortical dysrhythmia: a neurological and neuropsychiatric syndrome characterized by magnetoencephalography. Proc Natl Acad Sci USA. 1999;96(26):15222–7. DOI: 10.1073/pnas.96.26.15222

5. Moyer H, Mellett R, Vigneault K, McKeown M, Karlawish J, Augustine E, et al. Prevalence and impact of bypassing or overriding phase 2 trials in neurologic drug development. Neurology. 2024;103(1):e209533. DOI: 10.1212/WNL.0000000000209533

6. Wang F, Tian ZC, Ding H, Yang XJ, Wang FD, Ji RX, et al. A sensory-motor-sensory circuit underlies antinociception ignited by primary motor cortex in mice. Neuron. 2025;113(12):1947–68. DOI: 10.1016/j.neuron.2025.03.027

Part 1

Drug-resistant epilepsy affects roughly 30% of people living with epilepsy, and even when they go through surgery, fewer than 70% become seizure-free. For a long time, the working assumption behind that gap was imprecision: the surgical team didn’t find quite the right spot, or didn’t remove quite enough of it. A 2026 review by a team of neuroengineers at Lanzhou University makes a more uncomfortable case — that the gap isn’t mainly about precision at all. It’s about treating something that behaves like a network as if it were a single point.

The explanation for that gap goes back further than any AI paper — to 1988, and a discovery about something far smaller than a network: a single neuron.

The discovery that explains the gap

For most of the 20th century, the dominant explanation for brain disease was anatomical. Something was wrong because something was structurally wrong: a lesion, a malformation, a region of dead or damaged tissue. Find the damaged place, and you’d found the cause. Neurons themselves were understood mainly as relays — quiet until a signal arrived from elsewhere, then passing it along. The brain’s story was a story about places.

Rodolfo Llinás — one of the most influential neurophysiologists of the modern era — spent years recording directly from single neurons, in the cerebellum, the thalamus, the inferior olive, and found something that didn’t fit that story. Neurons aren’t quiet until triggered. Left alone, with no outside input at all, many of them generate their own continuous electrical rhythm: a constant, fast-changing voltage across the cell membrane, shifting from one moment to the next, driven by specific channels in the membrane that let charged ions in and out. That electrical change isn’t incidental to the neuron’s biochemistry — it is what drives it, opening and closing the channels that determine how the cell behaves next. Llinás laid this out in detail in a 1988 paper in Science, and it amounted to a different claim about what kind of object a neuron is: not a switch waiting for instructions, but a small, self-sustaining electrical generator, running on its own, every second, whether or not anything from outside ever reaches it.

That alone would have been a significant finding about single cells. What made it matter for medicine is the next step, which Llinás and colleagues took a decade later, published in 1999 in the Proceedings of the National Academy of Sciences. A neuron’s intrinsic rhythm doesn’t stay private. Neighboring neurons with the same kind of rhythm can lock into it together, oscillating in step, forming a group whose members are now doing something none of them could do alone. And a group’s rhythm can, in turn, lock together with a completely different group sitting somewhere else in the brain — a kind of long-distance electrical handshake between regions that aren’t anatomically the same place at all. Using magnetoencephalography to record this directly in patients, Llinás and colleagues found that several diseases that look nothing alike — chronic nerve pain, tinnitus, Parkinson’s disease, depression — shared the same broken version of this handshake: an abnormal slow rhythm in the thalamus, locked together with faster rhythms across the cortex, in a pattern they called thalamocortical dysrhythmia. The thalamus, in this picture, behaves something like a headquarters whose malfunction doesn’t stay local — it shows up, correlated, in distant cortical territory, producing a clinical picture that has nothing to do with where the “damage” supposedly is, because there often isn’t a localized lesion to find at all.

This is the real shift: disease stopped being only a question of which place was broken, and became a question of which relationship between places had broken down. A single neuron’s rhythm becomes a group’s rhythm, and a group’s rhythm becomes a conversation between groups in different parts of the brain and it’s at that last level, the conversation, that function and dysfunction actually live.

That’s the lens the rest of this story runs on. Epilepsy is simply where it has become a surgical decision rather than just a theory.

From one spot to a network

For most of the modern era of epilepsy surgery, the model was straightforward: find the spot where the seizure starts, remove it. The Lanzhou team’s 2026 review describes a paradigm shift away from that model, toward characterizing and modulating the dysfunctional brain network that actually generates and spreads the seizure, rather than hunting for one discrete epileptogenic zone, the clinical, surgically-actionable version of exactly what Llinás described decades earlier.

The evidence for this comes from signals that are measurable even between seizures, not only during them: brief, very fast oscillations in brain tissue, shifts in the balance between excitatory and inhibitory signaling, and patterns of connectivity between brain regions. The review traces a path from these local biomarkers, to mapping how they connect into a network, to using AI particularly deep learning models built to handle space and time together — to turn that network picture into something a surgical team can act on: where the network’s trouble is concentrated, and how likely a given patient is to do well after surgery.

What AI is actually good at right now

The case for taking this seriously isn’t just theoretical. A separate 2024 study tested an AI model built to interpret EEGs against a panel of human experts — not on the data it was trained on, but on a different patient population, recorded on different equipment, somewhere else entirely. The AI matched the experts almost exactly: 92% accuracy against 94%, with no meaningful difference across categories of abnormality (Mansilla et al., Epilepsia, 2024). That result matters because it shows a model can hold onto a real signal even when the recording conditions change completely — which is exactly the property a network-level read needs in order to be trusted outside the lab that built it.

Worth knowing, in plain terms: that AI tool isn’t a lab experiment, it’s a real product on the market today, sold through a company, and a couple of the people who built it also helped run the study testing it. That doesn’t make the result wrong — the actual test happened at an independent hospital using different equipment, which is the part that matters most — but it’s a detail worth knowing rather than skipping past.

Where it’s still going wrong

If Llinás was right that disease lives in the conversation between groups of neurons, then a model built to find that conversation should be the whole answer. It isn’t yet, and the Lanzhou review is candid about why — more candid than most papers in this space.

• First, incomplete sampling: most models are trained on data from a small number of centers, with a narrow slice of the real diversity of patients, electrode placements, and recording setups. A model that looks excellent on the data it grew up on can fail quietly the moment it meets a patient who doesn’t resemble that data.

• Second, unstable outcome labels: “surgical success” itself isn’t measured consistently from one study to the next. If the thing a model is being trained to predict is itself inconsistently defined, a high accuracy score can describe a moving target rather than a real clinical outcome.

• Third, limited interpretability: many of the best-performing models are black boxes. They produce a number or a probability without showing the clinician which features of the network drove that conclusion. That’s tolerable for a low-stakes recommendation. It’s much harder to accept when the decision on the other end is whether to operate on someone’s brain.

None of these are framed in the review as reasons to abandon the approach. They’re framed as the actual to-do list — the gap between a 38-year-old idea about how the brain works and a system a surgical team can responsibly act on.

Why this still matters

What AI does with epilepsy today is, in a very direct sense, the mathematical version of what Llinás found by hand. Where he and his colleagues identified a correlated rhythm between two specific regions using MEG, a deep learning model run across dozens of EEG or intracranial electrodes is doing the same kind of search at much larger scale — finding which groups of signals move together, which correlations hold across a whole network, and where in that network the trouble is concentrated. The math changed completely. The underlying claim about the brain — that the unit of disease is a network of correlated groups, not a single anatomical place — is the one Llinás made in 1988, and it’s the same one the Lanzhou review is building on today.

There’s a way to see this that goes beyond epilepsy. Picture the brain as a circle, and a single biomarker — one signal, one electrode, one timepoint — as a point sitting at its center. From that center, there are dozens of directions you could walk toward the edge, and each one reaches a different part of the circle. A model trained to follow one direction well still knows nothing about the other ninety-nine. That’s the actual difficulty of understanding the brain: not that any single signal is too faint to find, but that there are too many directions a single solution can take from the center, and no one direction explains the whole shape.

That’s the deeper lesson underneath both Llinás’s work and the Lanzhou review, and arguably underneath most of clinical AI right now. What made thalamocortical dysrhythmia matter wasn’t that Llinás found a new signal — plenty of people can find a signal. What mattered was that he built one coherent way of reading that signal which turned out to explain four diseases that had nothing else in common. That’s integration: not more data, not more discovery, but a paradigm general enough to extrapolate from one case to the next.

This may be the real gap in clinical neuroscience and the technology built on top of it — not a shortage of innovation, but a shortage of integration. The tools keep getting more powerful: more GPUs, larger models, more channels of data. None of that, by itself, gets closer to a paradigm that makes sense of what all that power is finding. A faster way to detect a signal isn’t the same as a better way to understand what it means, or where else that meaning applies. Whether the question in front of a clinician is a seizure, a tremor in Parkinson’s disease, the early signs of Alzheimer’s, or chronic pain that won’t respond to treatment, more computing power without that interpretive paradigm doesn’t bring you closer to a right answer. It just gets you to a wrong one faster.

The honest version of this story isn’t “AI is solving epilepsy.” It’s that AI has finally given researchers a way to act on something neurophysiology already knew — and that acting on it responsibly means building the paradigm first, fixing sampling, labels, and interpretability along the way, rather than mistaking more computing power for more understanding. A more interpretable model trained on more representative data, validated the way the 2024 EEG study was validated, is a more modest promise than “AI-powered epilepsy surgery.” It’s also the one that’s actually likely to reach a patient.

🔗 neuroedgenexus.com

References

• Wang Y, Yan Z, Gong Y, Lin W, Wang T, Liu Y, Han Y, Yang M, Liu M, Chen W, Hu B. Computational neuroelectrophysiology and artificial intelligence for drug-resistant epilepsy: recent advances, current challenges, and future directions. Journal of Neural Engineering. 2026;23(3):031003.

• Mansilla D, Tveit J, Aurlien H, et al. Generalizability of electroencephalographic interpretation using artificial intelligence: An external validation study. Epilepsia. 2024;65(10):3028–3037.

• Llinás RR. The intrinsic electrophysiological properties of mammalian neurons: insights into central nervous system function. Science. 1988;242:1654–1664.

• Llinás RR, Ribary U, Jeanmonod D, Kronberg E, Mitra PP. Thalamocortical dysrhythmia: A neurological and neuropsychiatric syndrome characterized by magnetoencephalography. PNAS. 1999;96(26):15222–15227.

Neuroscience & Alzheimer’s Prevention

SERIES EVOLUTION

Editions 21 → 22 → 23: From signals to biomarkers

Alzheimer’s disease does not begin with memory loss and the signal that precedes it may already be present every night, in the physiological record of how we sleep.

CLINICAL VIGNETTE

A 64-year-old neurologist presents for routine evaluation. Cognitive testing is normal. MRI is unremarkable. Blood biomarkers are within expected range. She reports no cognitive concerns.

Yet an overnight polysomnography recorded three years earlier shows a 22% reduction in slow-wave activity and increased sleep fragmentation compared to age-matched controls.

She has no diagnosis. No symptoms.

But the signal was there — recorded, timestamped, and not clinically interpreted.

Where this fits in NeuroEdge Nexus Series

Edition 21 established the body as data — demonstrating how digital biomarkers and wearable systems enable continuous physiological monitoring.

Edition 22 identified the translational bottleneck as systemic: governance, infrastructure, and implementation — not scientific discovery.

Edition 23 applies this framework to Alzheimer’s disease, where sleep may represent one of the earliest accessible physiological signals of preclinical pathology.

Beyond Symptoms: Where Alzheimer’s Begins

By the time cognitive symptoms emerge, neurodegenerative processes have often been evolving for years — if not decades.

This preclinical phase represents the primary window for intervention, yet remains largely inaccessible in routine clinical workflows.

The question is no longer conceptual, but operational:

How do we detect Alzheimer’s disease before symptoms — at scale, and in a way that health systems can act on?

Sleep is increasingly emerging as a candidate answer.

Study 1 — Sleep EEG as a Brain Health Biomarker at Scale

A study published in NEJM AI (2026) analysed 36,000 polysomnography recordings from 27,000 individuals across six cohorts.

Using deep learning, the model derived a Brain Age Index (BAI) directly from raw EEG, without expert-defined features.

The BAI showed significant associations with:

• cognitive performance

• dementia risk

• mortality

—including in individuals who were cognitively normal at baseline.

KEY FINDINGS (NEJM AI, 2026 | DOI: 10.1056/AIoa2500487)

• 36,000 PSG recordings across 6 cohorts

• AI-derived BAI from raw EEG

• Strong association with cognitive decline and dementia risk

• Detectable signal in preclinical individuals

The interpretation is significant. The study does not claim that sleep EEG diagnoses Alzheimer’s disease. It demonstrates that an AI-derived index of brain health — extracted from a single overnight EEG recording — carries meaningful prognostic information about cognitive decline and dementia risk at population scale. This is precisely the kind of scalable, non-invasive signal that the translational framework established in Edition 22 requires.

The biomarker is not created by the AI model. It is revealed through it

— extracted from a physiological record that already exists.

Study 2 — Wearable Sleep Recording: From Hospital to Home

The second key study in this edition demonstrates that the precision of sleep-based Alzheimer’s detection does not require a hospital-grade polysomnography suite. Published in npj Aging (2025), this research evaluated whether multimodal wearable sleep recordings could achieve clinically meaningful accuracy for Alzheimer’s disease screening.

Participants:

• 67 cognitively normal

• 35 Alzheimer’s patients

Using single-channel EEG + accelerometry and AI-driven analysis:

KEY FINDINGS (npj Aging, 2025)

• Detection accuracy: 0.90 (AD), 0.76 (prodromal AD)

• Wearable signals matched hospital PSG performance

• Physiological features outperformed sleep staging

This finding has direct translational relevance. It means the barrier to deployment is not technical — it is systemic. The signal is accessible from consumer-grade wearable devices. The analytical pipeline is available. This establishes that the signal is not confined to specialised laboratories.

It is accessible at the point of care — and potentially at population scale.

What We Are Actually Measuring During Sleep

Polysomnography captures a multi-system physiological state:

THE POLYSOMNOGRAPHY SIGNAL LANDSCAPE

• EEG — neural oscillations (slow waves, spindles, REM dynamics)

• ECG — autonomic regulation

• Respiration — ventilatory control and oxygenation

• EMG / movement — arousal and motor patterns

The signal is already available. What remains underdeveloped (as demonstrated by both studies in this edition) is the systematic interpretation of that signal in the context of neurodegenerative disease risk. The NEJM AI study shows what is possible at scale. The wearable study shows what is possible at the point of care. Together, they define the current frontier.

From Biomarker to Deployment Minimum Viable Pathway

• Step 1 — Risk Stratification
  Sleep EEG-derived indices integrated into at-risk populations (≥60, APOE4, subjective decline)

• Step 2 — Targeted Confirmation
  High-risk profiles trigger plasma, CSF, or PET biomarkers

• Step 3 — Longitudinal Monitoring
  Wearable sleep EEG enables continuous tracking over time

• Step 4 — Intervention Layer (Current State)
  Risk modification, sleep-targeted strategies, and trial inclusion

✅ Biomarker Strengths

• Non-invasive and repeatable

• Already recorded in clinical practice

• Detectable in preclinical individuals (NEJM AI)

• Accessible via wearable devices (npj Aging)

• Scalable to population-level screening

⚠️ Translation Barriers

• Not yet standardised for AD risk stratification

• No validated clinical intervention pathways

• Regulatory frameworks for AI-derived biomarkers evolving

• Reimbursement structures absent

• Long-term outcome data still limited

At present, no sleep-based biomarker meets regulatory criteria for standalone Alzheimer’s diagnosis or treatment initiation.

AI as an Enabling Layer — Not the Core

Both studies reviewed in this edition use artificial intelligence. In the NEJM AI study, deep learning extracts a brain health index from 36,000 recordings that no human expert could have derived manually. In the npj Aging study, AI-driven sleep staging enables wearable-grade recordings to match hospital-grade discriminative capacity.

In neither case does AI redefine the underlying physiology. The slow-wave activity is still there. The sleep fragmentation is still measurable. The amyloid-beta dynamics are still occurring. AI provides the analytical layer that makes these signals interpretable at scale — and that is precisely the role it should occupy.

AI does not create the biomarker. It reveals it — and makes it readable at a scale and resolution that changes what clinical practice can do.

Public Health and EU Policy Perspective

From a public health standpoint, the implications of these two studies are significant. Across Europe, Alzheimer’s disease represents a growing burden, with interventions still largely focused on symptomatic stages. The European Health Data Space (EHDS), established in March 2025, creates the regulatory architecture for health data sharing — including the kind of longitudinal sleep EEG data that this research depends on.

The AI Act establishes risk-based governance for AI systems used in healthcare — directly applicable to AI-derived biomarker tools like the BAI described in the NEJM AI study. As the European Brain Council has emphasised: without brain health, there is no health. The policy framework exists. The signal exists. The analytical tools exist. What is needed is coordinated deployment.

This is exactly the systems challenge that Edition 22 described — and which sleep-based Alzheimer’s biomarkers now make concrete: not a question of whether the science works, but whether the infrastructure surrounding it can be built with the same seriousness as the science itself.

System Responsibility — From Signal to Implementation

• Health systems → integrate sleep biomarkers into prevention pathways

• Regulators → define validation standards for AI-derived EEG biomarkers

• Payers → establish reimbursement models

• Industry → generate outcome data and standardise devices

The limiting factor is no longer signal detection

— it is coordinated system ownership.

The NeuroEdge Nexus Perspective

For decades, sleep has been recorded with high-resolution physiological tools.

The signal has been present.

What has been missing is:

• structured interpretation

• clinical integration

• system-level deployment

The evidence is now sufficient to define sleep as a candidate biomarker for early Alzheimer’s detection.

The remaining challenge is not scientific.

It is organisational.

EDITION 23 — KEY SCIENTIFIC POSITIONS

• Alzheimer’s pathology precedes clinical symptoms by years — early signal detection is the clinical priority

• Sleep EEG carries a Brain Age Index detectable across 36,000 recordings — Ganglberger et al., NEJM AI 2026

• Wearable single-channel EEG achieves 90% AD detection accuracy at home — npj Aging, 2025

• Sleep is a candidate biomarker — non-invasive, repeatable, and already embedded in clinical systems

• AI reveals the signal — it does not create it. The physiology is the biomarker; AI is the analytical layer

• EHDS + AI Act provide the regulatory foundation — clinical deployment requires coordinated systems investment

Scientific References

1. Ganglberger W, Sun H, Turley N, Tripathi A, Hadar P, Gupta A, Gallagher K, et al. Brain Health from Sleep EEG: A Multicohort, Deep Learning Biomarker for Cognition, Disease, and Mortality. NEJM AI. Published February 26, 2026. 2026;3(3). doi:10.1056/AIoa2500487

2. Wearable sleep recording augmented by artificial intelligence for Alzheimer’s disease screening. npj Aging. 2025. [Wearable EEG + accelerometry; AI-driven sleep staging via SeqSleepNet; n=102 (67 controls, 35 AD). Peer-reviewed publication. Full DOI pending confirmation — cited for methodological findings only.]

3. Lucey BP et al. Reduced non-rapid eye movement sleep is associated with tau pathology in early Alzheimer’s disease. Science Translational Medicine. 2019;11:eaau6550.

4. Winer JR et al. Sleep as a potential biomarker of tau and β-amyloid burden in the human brain. Journal of Neuroscience. 2019;39:6315–6324.

5. Mander BA et al. Sleep: a novel mechanistic pathway, biomarker, and treatment target in the pathology of Alzheimer’s disease. Trends in Neurosciences. 2016;39:552–566.

6. European Health Data Space (EHDS). Regulation (EU) 2025/327 of the European Parliament and of the Council. Entered into force March 2025.

NeuroEdge Nexus translates neuroscience, AI, and European regulatory frameworks into decision-grade strategic analysis. Season 2 (2026) focuses on governance, infrastructure coordination, and the implementation gap in digital brain health.