The Evo-Stress Blog

What prenatal stress research reveals about the coming challenge of physiological biotyping

---

A paper about maternal stress hormones and autism is almost guaranteed to become a story about whether stress during pregnancy causes autism.

That is not what the evidence shows. But a recent review connecting maternal stress physiology, placental regulation, and autism-related development is still worth reading closely — not for its conclusions, but because it exposes a problem that runs underneath the whole field. Stress science is getting much better at detecting patterned biological differences between people. It has not yet worked out how to say what those differences mean.

This will not be the last paper of its kind. Over the next few years, evidence connecting stressors, stress pathways, and stress mechanisms to autistic development and autistic health is going to arrive in volume — from prenatal cohorts, from immune and metabolic work, from mitochondrial studies, from wearable and biomarker datasets that did not exist five years ago. It will not arrive in an orderly sequence. Much of it will be underpowered, and a good deal of it will contradict the rest. Some of it will be the most affirming science autistic people have had: evidence that the exhaustion is real, measurable, and not a character failure. Some of it will be insulting, describing the same bodies as broken. A fair amount will manage both at once, in the same paper, sometimes in the same paragraph.

So it is worth being prepared. Waiting to see how each study turns out is not a strategy — the volume is too high and the interpretations will be settled long before the evidence is. What is needed is a way to read the wave as it arrives.

I am not going to name the review. That is deliberate. The difficulty I want to describe is not really its difficulty — it belongs to the field the review is working in, and naming it would turn a structural problem into a dispute about one paper and its authors. Nothing I say here is a finding about that team. It would apply to a dozen other papers I could have picked instead.

The gap between detecting a difference and interpreting it is about to matter more than it ever has, because stress science is moving toward classification: sorting people into groups based on coordinated autonomic, endocrine, immune, metabolic, and energetic profiles. Biotyping. And a classification system inherits whatever assumptions were built into it.

What the review is trying to do

The proposed pathway is straightforward enough to state:

• maternal psychosocial stress
• activation of the HPA axis
• placental regulation of glucocorticoid exposure
• altered fetal glucocorticoid conditions
• early differences in stress reactivity, attention, emotional regulation, and sensory processing

It is a plausible chain, and each link has some literature behind it. But two limitations sit at the center of it, and the review is reasonably candid about them. Direct evidence involving an actual autism diagnosis is thin. And most of the outcomes on the far end of the chain are not autism at all — they are broad developmental domains that are not specific to autism, or to any single diagnosis.

So the honest summary is: researchers are observing differences in how regulatory systems are organized and how they respond. They are not, in this literature, observing a mechanism that produces autism.

The real interpretive problem

Here is the pivot, and it is not a criticism of any one paper.

The problem is not whether physiology matters. It clearly does. The problem is deciding what a physiological difference means.

Any unusual regulatory pattern can reflect at least four different things:

1. Regulatory architecture — how the system is characteristically organized. Its structure, its gain settings, its coupling.
2. Current state — what the system happens to be doing today, under today’s demands.
3. Cumulative burden — the accumulated cost of sustained demand over time.
4. Reduced regulatory bandwidth — a constrained ability to track demand, mount an adequate response, and recover across the contexts that matter.

These are not shades of the same thing. They call for entirely different responses. A system that is organized differently and working well needs a fitting environment. A system that is temporarily loaded needs recovery. A system carrying years of accumulated cost needs that cost addressed. A system running out of adaptive margin needs the margin restored, urgently.

A different architecture can also be vulnerable, medically complex, or unsustainable under particular conditions. The point is not to presume either health or pathology from difference alone.

A cross-sectional physiological profile usually cannot distinguish these reliably. One blood draw, one cortisol curve, one clustering analysis at a single time point may collapse all four into a profile that sits some distance from the group average. The distance is real. What produced it is not visible in the measurement.

Applying that to prenatal research

This is where the distinction earns its keep.

Elevated or flattened cortisol does not identify one biological story. It may reflect a proportionate response to genuine current demand. It may reflect a chronically altered state. It may reflect reduced recovery capacity. It may reflect different circadian organization, metabolic constraint, medication, sleep disruption, or a stable feature of that person’s regulatory architecture that has been there since long before the pregnancy.

The same is true downstream. A child’s sensory reactivity, social attention, or stress response may represent neurodevelopmental organization, a transient state, an overloaded system, or an interaction among all three at once.

Which means researchers cannot move directly from “different physiology” to “mechanism of autism.” The step from difference to mechanism requires estimating how much of the observed profile reflects architecture, present state, accumulated burden, and remaining capacity — and standard designs rarely permit that decomposition. These are analytically distinct and empirically entangled; in most people, several are operating at once.

Compared with whom?

There is a second question underneath the first, and it is the one I think matters most.

When a profile is called dysregulated, dysregulated relative to what?

Every classification system needs a reference class — a set of people whose physiology defines the normal, well-regulated configuration against which everyone else is scored. In stress science, those reference samples are usually general-population cohorts, aging cohorts, or disease cohorts. In all of them, neurodevelopmental variation is typically absent, underrepresented, actively excluded by screening criteria, or statistically controlled away as a nuisance variable.

The consequence is quiet and enormous: neurotypical regulatory patterns become the implicit physiological norm. Not by argument, and not by anyone’s intent. By sampling.

Applied to pregnancy research, this turns into a set of questions almost nobody is asking:

• What does a “normal” maternal cortisol rhythm mean across different neurotypes?
• Are autistic and ADHD pregnant people adequately represented in the samples that define it?
• Does sensory overload produce a different endocrine pattern that remains functional in context, or one that reflects accumulating cost?
• Are sleep, masking, autonomic differences, medication, and environmental mismatch measured — or left out and absorbed into the error term?
• Is a fetal or infant profile classified as abnormal because it is harmful, or simply because it is uncommon in the reference sample?

Uncommon and harmful are not the same finding. A model built on a sample that excludes a population is predisposed to classify that population’s recurring differences as deviations from the norm. That may reveal a real difference. It does not establish dysfunction.

This is also why the question is who decides rather than when. There is no point in the physiology where difference tips over into disorder — no threshold sitting in the cortisol curve waiting to be located. There is a reference class, a set of people who were or were not in it, and a judgment about what departing from it means. Those are choices, made by researchers, mostly at the design stage, mostly without anyone noticing they were choices at all.

Why this could reproduce maternal blame

This section needs saying carefully, because the risk is real and it arrives dressed as rigor.

A linear model that begins with “maternal stress” places the pregnant person at the head of the causal chain. She becomes the source of biological risk. But the stress in question is frequently produced somewhere else entirely — by poverty, discrimination, inaccessible or dismissive healthcare, hostile sensory environments, violence, caregiving strain, or simply the absence of adequate support. The physiology is hers. The demand very often is not.

Now add the neurodiversity layer. If an autistic pregnant person’s regulatory profile differs from a neurotypical reference class, that difference can be read as pathology before anyone has determined whether it reflects architecture, current overload, environmental mismatch, or genuinely reduced capacity. Her body becomes evidence of risk she is presumed to be transmitting.

This is how the pathology paradigm moves out of behavior and into physiology — and it will look more legitimate there, because it will be quantitative.

What better biotyping would ask

The alternative is not to stop measuring. It is to ask a different set of questions of the same measurements:

• How is this system organized?
• What demands is it currently meeting?
• How much adaptive range remains?
• How quickly does it recover?
• How does the environment change the profile?
• Is the pattern stable across different challenges and contexts, or does it appear only under some?
• Does it predict sustainable functioning — or only distance from the sample average?

That last one is the whole argument in a sentence. A profile that predicts sustainable function is a clinically meaningful finding. A profile that predicts distance from an average is a description of the sample.

Biotyping done this way generates architectural hypotheses — claims about how a system is built that can be tested against how it behaves under multiple perturbations. Biotyping done the other way ranks people against a supposedly universal healthy configuration that was never universal to begin with.

What a useful prenatal study would look like

Concretely: a study that repeatedly measures maternal sleep, lived conditions, cortisol rhythms, autonomic regulation, nutrition, immune and metabolic state, and placental function across the pregnancy rather than at one point. That characterizes parental and fetal neurodevelopmental architecture instead of treating it as a confound. That then follows children across several outcomes — not only autism diagnosis, but sleep, sensory regulation, autonomic recovery, pain, communication, adaptive functioning, and medical burden.

It would also treat the placenta as an active regulatory interface — transforming endocrine, immune, metabolic, vascular, and nutritional signals — rather than as a passive barrier that either succeeds or fails at blocking maternal cortisol. A barrier can only be intact or breached. An interface has its own organization, and can be well or poorly matched to what it is regulating.

And a prediction worth testing: prenatal regulatory patterns may turn out to be far more predictive of later regulatory burden than of autism itself. If so, the field has been looking for the wrong outcome — and has been looking past the one that would actually help people.

The larger argument

This is the case I develop at length in Stress Biotyping and the Neurodiversity Paradigm: Toward a Mechanistic Account of Neurodivergent Regulatory Architecture.

Stress science is moving quickly toward classification based on coordinated autonomic, endocrine, immune, metabolic, and energetic profiles. Much of the work driving that shift is exactly the right kind. Naviaux’s cell danger response, Picard’s mitochondrial and energetic accounts, and the allostatic-load tradition running back through McEwen all treat physiology as a dynamic system with states, sequences, and costs rather than as a list of biomarkers. My concern is narrower than any of them, and it attaches to a specific step: what happens when that work is operationalized as classification, when profiles become clusters and clusters become categories that people are sorted into. That step could produce genuinely better, more individualized science — physiology-first, less dependent on behavioral checklists, more able to explain why the same demand costs one person far more than another. But unless neurodevelopmental variation is written into its foundations, it may simply reconstruct the pathology paradigm at the level of the body.

A related case: mitochondrial allostatic load

Mahony and O’Ryan (2022) propose mitochondrial allostatic load as a mechanism linking chronic stress — early-life stress, and the sustained demand of camouflaging — to depression, suicidality, and burnout in autistic people. Their framing is already neurodiversity-aligned: the stated target is internal distress and unmet need, not autistic behavior.

What the ESF arc adds is the interpretive layer. Toward an Emergent Paradigm for Neurodiversity and Health situates mitochondrial change inside a coupled brain–body–environment system, where the same finding may reflect neurotype architecture, current regulatory state, or accumulated demand — rather than defaulting to damage. Regulatory Bandwidth then supplies the dynamical account: how sustained compensatory demand can end in a nonlinear loss of accessible capacity rather than a gradual decline.

The positive criteria are straightforward:

• treat neurodevelopmental variation as initial conditions, not noise;
• distinguish architecture from current state, cumulative load, and present capacity;
• measure across multiple perturbations and contexts, since architecture cannot be read from a single response;
• define outcomes through sustainability, fit, recovery, and participatory well-being;
• intervene to restore adaptive margin rather than to normalize architecture.

The paper develops the methodological consequences that a blog post can only gesture at: how reference classes should be constructed, why biotypes require repeated perturbations rather than static clusters, and how sustainable function, recovery, and environmental fit can replace distance from an average as the relevant validation criteria.

None of this denies illness, disability, or medical need. Some regulatory profiles genuinely reflect suffering that should be treated. The claim is narrower and, I think, harder to argue with: you cannot know which profiles those are until you can tell a differently organized system from a depleted one.

The interpretive layer

The history of medicine is full of stretches where measurement outran interpretation. The microscope, the X-ray, the electrocardiogram, the MRI — each made invisible biology visible, and each was followed by a period in which visibility got mistaken for understanding. Seeing a thing does not tell you what it is doing there.

Physiological biotyping is entering that stretch now. It will become very good at detecting coordinated biological differences. The harder task is deciding whether a particular difference reflects architecture, adaptation, accumulated cost, constrained capacity, or some combination of them — and that decision is not contained in the data. Biomarkers do not arrive labeled. They become architecture, or adaptation, or damage, only inside a model, and the model is supplied by us.

None of this is unique to stress physiology, or to autism. Metabolomics, epigenetics, immune phenotyping, connectomics, and every clustering algorithm turned loose on a biobank face the same problem: a measured difference is not yet a finding. Better measurement without better interpretation does not correct old assumptions. It renders them at higher resolution.

The warning

The next generation of autism science may not describe neurodivergent people as behaviorally defective. It may describe their autonomic, endocrine, immune, metabolic, and placental profiles as dysregulated instead.

That will sound more precise. It may even be called personalized medicine.

But unless we ask how the reference profile was built, whether neurodevelopmental variation shaped the model, and whether difference has been separated from depleted capacity, the underlying logic will not have changed at all. It will only have moved somewhere harder to see and harder to contest.

The problem is not that stress science is entering the body. The problem is that the neurodiversity paradigm has not yet entered stress science deeply enough.

---

This article introduces the argument developed more fully in “Stress Biotyping and the Neurodiversity Paradigm: Toward a Mechanistic Account of Neurodivergent Regulatory Architecture,” currently under review at Neurodiversity. The paper examines how physiological biotyping can distinguish regulatory architecture from current state, accumulated burden, and constrained capacity — without making neurotypical physiology the default standard of health.