Lab finding — week of May 25 2026 This is the conceptual closing session of Lua Labs' L2 line (HPA-HPO axis). Sub-topic 2.5 — integrative digital biomarkers of the HPA axis.

Carmen returns, now with 60 days of follow-up

Carmen is 47. Her labs are still "normal": borderline FSH, estradiol within range, clean thyroid profile. Her endocrinologist congratulated her and told her to come back in six months.

But after 60 days of recording how she feels each morning, how she eats, how she sleeps and when each symptom appears, the full picture tells a different story. Consecutive days with morning energy at 2 out of 5. Mood peaks in the luteal phase that no longer resemble the ones she had two years ago. A very clear pattern of caffeine after noon followed by sleep-onset insomnia. Two weeks without nopal or a single fermented food. And a sleep fragmentation score that her Apple Watch reports in the 80th percentile of her own personal history.

Each of those data points, alone, means nothing. Together they paint a picture of perimenopausal hormonal allostatic load that her serum profile cannot paint — because it was not designed to paint it.

This week's question at Lua Labs is: can we make that picture operational? Can we build, from the most recent scientific literature, a score that integrates the multiple dimensions of the stress axis and makes sense both to Carmen and to her doctor?

Allostatic load: the construct endocrinology never fully digitized

The term allostatic load was introduced by Bruce McEwen and Teresa Seeman in the mid-nineties to name something physiology had intuited for decades: the body is not damaged by acute stress, but by the accumulated wear of repeatedly adapting.

When something stresses you — a difficult call, a night of poor sleep, a cramp, a financial worry — your body activates a coordinated response: cortisol rises, heart rate rises, blood pressure rises, the immune system reorganizes. This is called allostasis: literally, "stability through change." It is adaptive and useful.

The problem appears when that activation stops turning off cleanly. When cortisol does not return to its nighttime nadir. When heart rate during sleep loses amplitude. When the immune system stays in low-grade inflammatory mode. When the microbiome loses the diversity that used to buffer those signals. That accumulated structural wear is allostatic load.

McEwen and Seeman proposed measuring it with a panel of serum markers: cortisol, DHEA-S, IL-6, fibrinogen, HDL, blood pressure, HbA1c, abdominal circumference. The Allostatic Load Index (ALI) counts how many of those markers are in the risk quartile. It works — it predicts mortality, cognitive decline and cardiometabolic risk. But it requires venous blood draw, laboratory analysis and a standardized protocol.

Thirty years later, the question the literature is beginning to answer seriously is: can allostatic load be translated into the digital domain? Can it be measured with a wearable, a food log and a record of how you feel — without blood?

The paper that changed the conversation in 2025

In 2025, the first study appeared that crossed both things in the same cohort. Wedge and collaborators, published in American Journal of Physiology — Regulatory, Integrative and Comparative Physiology, measured simultaneously:

The classic serum Allostatic Load Index (full multibiomarker panel)
Continuous military-grade wearable data (photoplethysmography + accelerometry)

…in a cohort undergoing arduous military training. The central result:

Elevated allostatic load has a specific digital signature: chronically elevated and variable cardiometabolic activity during the day, combined with attenuated variation in heart rate during sleep.

Three lessons reorganized how we think about the problem:

First. The signature of allostatic load appears during sleep, not during the day. Daytime heart rate is contaminated by activity, food and emotions. Heart rate during sleep is cleaner — and that is where structural wear becomes visible.

Second. Allostatic load does not show up as a high or low value. It shows up as instability — abnormally high day-to-day variability. This changes how the data must be read: the average is not enough; dispersion matters.

Third. The paper is explicit: the digital phenotype does not replace clinical measurement. It is a proxy for early detection and trend monitoring. This matters because it avoids the simplistic language of "your watch tells you if you have burnout."

The counterintuitive paper: adding cortisol does not improve prediction

Another 2025 finding redirects the strategy. Kim and collaborators (published in Digital Health) trained machine learning models to classify fatigue using:

HRV variables only (heart rate variability)
HRV + salivary cortisol

The HRV-only model reached AUC = 0.774. The HRV + salivary cortisol model reached AUC = 0.741 — meaning cortisol did not add additional information and in fact marginally reduced performance, probably because of the inherent noise in its measurement.

The implication is deep and challenges the natural reflex to "measure cortisol with more home kits." If digital proxies (HRV, sleep, subjective fatigue, behavioral patterns) already capture enough information from the stress axis to predict clinically relevant outcomes, then the priority is not adding more biochemical measurements — it is building better digital instrumentation of behavior.

Fragmented sleep disrupts cortisol by itself

One of the clearest mechanistic papers available today for understanding what happens specifically in perimenopause is Grant and collaborators (2023, Journal of Clinical Endocrinology & Metabolism). The design is exquisite: they took 22 healthy premenopausal women and exposed them to an experimental protocol of sleep fragmentation + pharmacological suppression of estradiol with leuprolide, to create a controlled model of menopause.

Result:

+27% bedtime cortisol (loss of the nighttime nadir)
−57% in morning awakening cortisol (attenuated CAR)
The effect persisted under estradiol suppression

The operational conclusion is important: fragmented sleep disrupts the stress axis by itself, not as a secondary consequence of hormonal change. In natural perimenopause — where nighttime awakenings from hot flashes or anxiety are common — this mechanism accumulates night after night, week after week.

That is why any attempt to build a digital score of hormonal allostatic load has to include sleep fragmentation as its own dimension, not as a simple downstream symptom.

The cohort that showed self-report is enough to find phenotypes

Another 2025 paper — Tariyal and collaborators in npj Women's Health — analyzed 147,501 symptom logs from 4,789 users of the MenoLife app. They applied hierarchical clustering, K-means on principal components and binomial network analysis. They found three clear clusters: pre-, peri- and menopausal — with correlated sub-phenotypes of mood, cognition, skin, digestion, nervous system and sexuality.

The key methodological point: the phenotypes emerged from purely self-reported data. No wearable, no cortisol, no AMH. Only symptoms logged consistently.

This shows something the clinical conversation still has not fully absorbed: the depth and granularity of behavioral check-ins are enough to distinguish valid hormonal phenotypes. Biochemical measurement is valuable, but it is not a prerequisite for doing useful science about women's bodies.

Meal timing modulates cortisol rhythm

Two papers close the picture of what a digital score can capture without blood. Loy and collaborators (2023, MY-CARE cohort) found that skipping breakfast is associated with lower morning cortisol, and that a late-meal pattern alters the coordinated rhythm of melatonin and cortisol. Pickering and collaborators (2025, Nutrients) review the mechanisms: the macronutrient composition of breakfast (protein + tyrosine), timing relative to the natural cortisol peak, the duration of the feeding window — all modulate the amplitude of cortisol rhythm through the circadian clock system.

The practical implication is direct: the time you first eat during the day is a stress-axis variable, capturable without any additional hardware. And almost nobody tracks it.

What a serum test does not capture — the six dimensions

Combining the above, the central conceptual idea of this session emerges: the stress axis has at least six partially independent measurable dimensions, and most standard clinical evaluations capture only one or two.

Dimension 1 — Biological buffer. Gut microbiome diversity and the density of prebiotics in the diet. It measures the system's buffering capacity. It is built over a 14-day window from a structured food log.

Dimension 2 — Autonomic tone. Vagal balance measured through sleep, alcohol, food and, optionally, wearable HRV. It measures the parasympathetic counterweight that turns off the stress response.

Dimension 3 — Active allostatic load. Subjective stress (PSS-4), caregiving load for dependents (a dominant predictor in Latin American cohorts), luteal-phase specific symptoms. It measures the accumulated input to the system.

Dimension 4 — Diurnal cortisol phenotype. The shape of the curve — high reactive amplitude versus flat collapsed — digitally derivable from patterns of morning and evening energy, type of insomnia, and the presence and timing of hot flashes. It is the dimension that most changes how the other five should be read.

Dimension 5 — Ovarian load. The ovarian paracrine CRH axis, capturable through cycle regularity, androgenic symptoms and concurrent stress. Specific to women with active ovaries.

Dimension 6 — Circadian calibration. Time of first food, sunlight exposure in the first hour of the day, timing of first caffeine. It is the most modifiable dimension in the short term — and almost never measured.

Each dimension operates on a different temporal scale. The biological buffer moves over weeks to months. Autonomic tone over days to weeks. The cortisol phenotype is relatively stable as a biotype. Circadian calibration changes day to day. That difference in scales is not a problem — it is a strength: it allows an integrated score to capture signals that a single biomarker would never capture. The "slow drift" of the microbial buffer can precede the "fast decompensation" of the cortisol phenotype by weeks.

Why this matters especially in perimenopause

In perimenopause, two hormonal systems enter transition at the same time. The ovary begins to falter — intermittent anovulation, shortened luteal phase, erratic estradiol oscillations — right when the stress axis has already spent years accumulating wear. Progesterone, which in younger women buffers cortisol by competing for the glucocorticoid receptor, becomes intermittent. Estradiol, which normally modulates the sensitivity of the HPA axis, oscillates violently.

The result: a perimenopausal woman can have substantially higher hormonal allostatic load than her age suggests, without any standard test reflecting it. Her FSH says "almost." Her estradiol says "normal." Her thyroid is clean. And still, her body is signaling clearly — just through variables that a 15-minute visit does not capture.

This is what Carmen experiences when she is told to come back in six months. It is what the public conversation about menopause still has not fully articulated. And it is why the idea of an integrative digital score, computed from what a woman already does — recording how she feels, what she eats and how she sleeps — is an intellectually relevant piece.

How Lua Labs is thinking about this

At Lua's lab this week, we are crossing what is being published in the digital allostatic load literature with the mechanistic reasoning built in previous sessions. The synthesis is the concept of the Integrative Digital Biomarker of the HPA axis (BDI-HPA): a conceptual architecture with six dimensions, a modulating phenotype (the cortisol biotype described in the previous session), a composite formula with weights derived from the literature, and a validation protocol with defined outcomes.

The BDI-HPA is not a cortisol measurement. It is not a diagnosis. It does not replace the clinical serum panel for allostatic load. Its conceptual utility is as a trend and self-knowledge score — a different object, with its own use: early trend detection, phenotype segmentation, personalized alerts for decompensation.

Three important clarifications about what this means today:

The BDI-HPA is currently a literature-derived conceptual proposal, not a feature already available in the app. This note documents the public formulation of the concept, not its deployment.
The hypothesis the formulation leaves explicitly open — and that will be evaluated over time — is whether a six-dimensional composite, derived only from behavioral and food-log variables, can predict perimenopausal decompensation with AUC ≥ 0.70, equivalent to or better than a null model based only on subjective stress + age + cycle phase.
Any translation of the concept into product will require rigorous validation, objective criteria for decompensation, and a clear ethical line: the BDI-HPA should never be presented as "your estimated cortisol" — only as a trend index with self-knowledge utility.

What changes after this session

Lua Labs' L2 line — the HPA-HPO axis in women — closes conceptually this week with the idea that the six biomarkers built across previous sessions are not six separate things, but six dimensions of the same system. Perimenopausal hormonal allostatic load is not a value; it is a profile — a vector with six axes that changes from week to week and must be read in its full shape.

For Carmen, the concrete point is that the clinical language she needs for her next visit is no longer "I am stressed" or "I have hot flashes" or "I am not sleeping well." It is: "my diurnal energy curve changed over the past six weeks, my sleep is more fragmented, my feeding window shifted, and my mood in the luteal phase no longer looks like last year's. My test says 'normal.' My trajectory says 'this is moving.'"

That is the difference between a test and a phenotype. And that is why the category of longitudinal hormonal intelligence exists — because the conversation women in perimenopause deserve to have with their doctors does not fit in a single number, or a single day, or a single test.

References cited

Wedge FM, Phillips MR, Brown PJ, Friedl KE, Hoyt RW, Buller MJ, et al. (2025). Identifying a digital phenotype of allostatic load: association between allostatic load index score and wearable physiological response during military training. American Journal of Physiology — Regulatory, Integrative and Comparative Physiology. DOI: 10.1152/ajpregu.00216.2025

Kim JE, Kim NH, Choi SK, Lee J-Y, Lee K, Han JS. (2025). Machine learning-based fatigue classification using heart rate variability and cortisol: A multimodal approach to wearable health monitoring. Digital Health. DOI: 10.1177/20552076251395570

Grant LK, Coborn JE, Cohn A, Nathan MD, Scheer FAJL, Klerman EB, et al. (2023). Effects of Sleep Fragmentation and Estradiol Decline on Cortisol in a Human Experimental Model of Menopause. Journal of Clinical Endocrinology & Metabolism, 108(11): e1347–e1357. DOI: 10.1210/clinem/dgad285

Velazquez Sanchez C, Dalley JW. (2025). The cortisol awakening response: Fact or fiction? Brain and Neuroscience Advances, 9. DOI: 10.1177/23982128251327712

Tariyal R, Yoo S, Kennedy C, Kraschnewski JL, et al. (2025). Clustering of >145,000 symptom logs reveals distinct pre-, peri-, and menopausal phenotypes. npj Women's Health. PMC11699220

Loy SL, Cheng TS, Colega M, Cheung YB, Godfrey KM, Tan KH, et al. (2023). Chrononutrition is associated with melatonin and cortisol rhythm during pregnancy: Findings from MY-CARE cohort study. Frontiers in Nutrition. PMC9852999

Pickering G, Mazur A, Trousselard M, et al. (2025). 'Feeding the Rhythm' — Effects of Food and Nutrients on Daily Cortisol Secretion: From Molecular Mechanisms to Clinical Impact. Nutrients. PMC12653711

Crook H, Ramirez A, Hosseini AA, et al. (2023). Towards a consensus definition of allostatic load: a multi-cohort, multi-system, multi-biomarker individual participant data meta-analysis. eBioMedicine. PMID: 37100008