Longitudinal HPA-axis phenotypes in female allostatic load

Date: 2026-05-25 Researcher: Lua Labs — Scientist Classification: Applied AI / Neuroendocrine / Metabolomics Line: L2 — HPA-HPO Axis Subtopic: 2.5 — Longitudinal HPA-axis phenotypes (integrative session)

External sources

1. 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.

2. 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 (SAGE). DOI: 10.1177/20552076251395570. PMC12602915.

3. 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. PMID: 37207451. PMC10584010.

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

5. 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.

6. 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. (cited as mechanistic evidence for meal-timing × cortisol-rhythm in women; analogous applicability to perimenopause)

7. 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.

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

Framing

Isolated physiological and behavioral data rarely suffice to describe hormonal allostatic load. Their value appears when they are interpreted as trajectories: sleep, circadian rhythm, perceived stress, autonomic tone, metabolism and reproductive context change on different time scales, but converge on the same HPA-HPO axis.

The question for L2.5 is whether a composite phenotype can describe that convergence without replacing clinical cortisol measurements or promising individual diagnosis. The scientific objective is to separate a real longitudinal signal from point noise, self-report bias and inter-individual variability.

Base knowledge (what I know before searching)

Why a composite, not a single biomarker

The HPA axis is not a single variable. It is a multilevel system with at least five partially independent measurable dimensions:

1. Input load — how much chronic stress the system receives (PSS-4, caregiving burden, night shifts, life events)
2. Circadian architecture of cortisol — diurnal amplitude, CAR, nocturnal nadir (L2.4 phenotype A vs B)
3. Autonomic tone — HRV (RMSSD, LF/HF), respiration, resting heart rate. Vagal tone is the parasympathetic counterweight to the HPA axis
4. Metabolic resilience to stress — consolidated sleep, feeding window aligned with AM cortisol, absence of post-12pm caffeine, morning sunlight exposure
5. Endogenous biological buffer — microbiome (estrobolome + progesterobolome), SCFA-mediated vagal afference, compensatory DHEA-S

The categorical error repeated in the clinical literature is taking one of these levels (typically a single-point salivary cortisol measurement) and pretending it captures the system. The allostatic load literature (McEwen, Seeman, Crook 2023) has argued the opposite for 30 years: AL is composite by definition. But that same literature has measured AL with invasive serum markers (cortisol, DHEA-S, IL-6, fibrinogen, HDL, blood pressure, HbA1c, waist circumference), accessible only in cohorts with periodic extraction protocols.

The scientific opportunity is to translate that construct into observable longitudinal phenotypes without confusing them with clinical diagnosis: sleep patterns, circadian rhythm, subjective stress, autonomic tone, symptoms and metabolic context can act as trend signals when analyzed together.

What a longitudinal composite phenotype contributes

A composite phenotype should not be interpreted as a universal scale or as a substitute for serum cortisol. Its potential utility is describing trajectories: when a system becomes more variable, when it loses circadian amplitude, when sleep fragmentation precedes symptoms and when reproductive context changes the expected direction of a signal.

Three critical observations:

1. The cortisol phenotype may be structural, not continuous. A curve with high CAR and a flat curve are not merely different intensities of the same phenomenon; they may represent different physiological architectures.

2. Temporal granularities are not the same. Diet, sleep, stress, reproductive phase and autonomic tone change across different windows. A scientific model must respect those scales and avoid adding them as if they were equivalent.

3. Validation must separate physiological signal from measurement noise. Self-reports, wearables, salivary cortisol and serum biomarkers respond to different sources of error; external triangulation is required before individual inference.

What I know about digital composite biomarkers — the state of the art

Three principles established in the operational allostatic load literature (McEwen-Seeman MacArthur framework, refined by Crook 2023):

1. Normalization by quartiles, not raw z-scores. AL components are not Gaussian. The sum is made by counting how many components are in the risk quartile (typically the worst quartile of the population distribution). This prevents a single extreme component from dominating the score.

2. Weights derived from literature, not from initial regression. Until there is a validated clinical outcome, assigning weights by regression is overfitting. Better: assign weights based on the magnitude of mechanistic evidence (weight 1 = reference, weight 2 = RCT evidence, weight 3 = causal mechanistic evidence).

3. Latent dimensions, not a flat sum. Classic allostatic load has four partially independent dimensions (cardiovascular, metabolic, inflammatory, neuroendocrine). Collapsing them into a single number loses clinically relevant information: a composite phenotype is better interpreted as a vector of dimensions than as a single scalar.

Findings from recent papers

Finding 1 — The first validated association between classic serum AL and a wearable digital phenotype (Wedge et al. 2025, Am J Physiol)

The paradigmatic 2025 study: for the first time, a cohort (arduous military training, n not extracted but ≥40 participants according to the standard method for these protocols) had simultaneous measurement of classic allostatic load index (serum multibiomarker panel) + continuous military-grade wearable data (photoplethysmography PPG + triaxial accelerometry). The central result: elevated AL is associated with a specific digital phenotype — chronically elevated and variable cardiometabolic activity + attenuated variation in heart rate during sleep.

Two scientific lessons follow. First, the AL signature appears mostly during sleep — nocturnal heart-rate variation, not daytime values contaminated by activity. Second, AL manifests in variability and instability, not in mean values: the associated phenotype is not "high rate" but "unstable rate". Importantly, the authors frame the digital phenotype as a complement for early detection and trend monitoring, not as a substitute for serum measurement.

Finding 2 — Adding cortisol to HRV does NOT significantly improve fatigue prediction (Kim JE et al. 2025, Digital Health)

Counterintuitive 2025 finding: an ML model using only HRV variables reached AUC = 0.774 for fatigue classification; after adding salivary cortisol, AUC changed to 0.741 — that is, cortisol did not contribute significantly and in fact marginally reduced performance (probably because of measurement noise).

Finding 3 — Sleep fragmentation increases bedtime cortisol by 27% and lowers CAR by 57% in an experimental menopause model (Grant et al. 2023, JCEM)

This is the clearest mechanistic paper available to support the inclusion of sleep fragmentation as a component of a composite HPA-load phenotype. In a human experimental protocol (n=22 healthy premenopausal women subjected to sleep fragmentation + E2 suppression with leuprolide to mimic menopause), sleep fragmentation produced:

• +27% bedtime cortisol (elevated evening cortisol = loss of nadir)
• −57% CAR (attenuated morning peak)
• WASO (wake after sleep onset) correlates positively with bedtime cortisol and negatively with CAR

Combined with experimental E2 decline, the effect on cortisol remained — that is, sleep fragmentation is a disruptor of the HPA axis independent of estrogenic state. This directly validates:

• The sleep component within autonomic resilience
• The "type of insomnia" component within the diurnal cortisol phenotype (maintenance insomnia → fragmentation → emerging phenotype B)
• The inclusion of WASO or number of awakenings (captured by wearable) as a variable in a composite HPA-load phenotype

Critical: this is a protocol in premenopausal women with artificially suppressed E2 — it is not natural perimenopause. But the mechanism (fragmentation → bedtime cortisol up + CAR down) is probably generalizable, and Grub 2021 (Swiss Perimenopause Study, cited in L2.4) confirms the cortisol-up pattern during real perimenopause.

Finding 4 — CAR is real but more heterogeneous than assumed (Velazquez Sanchez & Dalley 2025, Brain & Neuroscience Advances)

Critical 2025 review on the cortisol awakening response. Conclusion: CAR exists (it is not an artifact), but its magnitude is highly individual and modulated by factors that single-day measurements do not capture. A single CAR measurement is noisy. Clinical utility comes from the longitudinal trajectory, not from the point value.

Finding 5 — Clustering of 145k symptom logs reveals 3 digitally discriminable hormonal phenotypes (Tariyal et al. 2025, npj Women's Health)

A cohort of 4,789 women with 147,501 symptom logs was analyzed with hierarchical clustering + K-means over principal components and binomial network analysis, identifying three clear clusters — pre-, peri- and menopausal — with correlated sub-phenotypes across mood, cognition, skin, digestion, nervous system and sexual domains.

The most methodologically relevant point: the phenotypes emerged from purely self-reported data, without wearables, without cortisol and without AMH. It is proof of concept that well-structured longitudinal symptoms suffice to discriminate valid hormonal phenotypes. The limitation: the work describes clusters but does not build an interpretable or predictive composite phenotype from them — the step from description to actionable signal remains an open area in the literature.

Finding 6 — Meal timing modulates the diurnal rhythm of cortisol (Loy 2023 MY-CARE; Pickering 2025 "Feeding the Rhythm")

Two papers that justify including chrononutrition as a dimension of a composite HPA-load phenotype. MY-CARE (pregnancy cohort, n=535) found that skipping breakfast is associated with lower morning cortisol and that a late-meal pattern alters the diurnal rhythm of melatonin-cortisol. Pickering 2025 (Nutrients) reviews the mechanisms: breakfast macronutrient composition (protein + tyrosine), timing relative to the cortisol peak, and feeding window all modulate the amplitude of the cortisol rhythm through SCN-clock genes-HPA.

Complete molecular/endocrine mechanism — proposed physiological system

FROM BELOW (input):
                                        Chronic cortisol → GR without P4 → dominant signaling
PSS-4 (subjective stress) ─────┐                              ↓
Caregiving burden ─────────────┤    CRH-PVN ──→ ACTH ──→ Adrenal cortisol
Night shifts ──────────────────┤      ↑               (diurnal curve)
Life events ───────────────────┤      │                              ↓
                              │      │           ┌────────────────────────────┐
                              │      │      Phenotype A          Phenotype B
                              │      │      High CAR             Flat CAR
                              │      │      amp. preserved       amp. collapsed
                              │      │      ↓                            ↓
                              │      │   Anxiety/insomnia       Fatigue/brain fog
                              │      │   sleep onset            subclinical depression
                              ▼      │   AM hot flashes          diffuse hot flashes
                       Hypothalamic microglia
                       ↑    (afferent vagal input)
                       │              ↑
Dysbiosis ─────────────┤              │
Low-grade LPS ─────────┤         Vagal NTS axis ←─── enterochromaffin cells ←── SCFA
Hypothalamic TLR4 ─────┘              ↑                                       ↑
                                      │                                       │
                                Efferent vagal tone ───────────────────  Microbiome
                                      ↑                                  (estrobolome,
                                Consolidated sleep                      progesterobolome,
                                Breathing 4-6 cpm                       neurobolome)
                                AM sunlight                                  ↑
                                                                              │
                                                                       Dietary diversity
                                                                       Fermented foods
                                                                       Fermentable fiber

Integrated physiological reading:

New dimensions that recent literature makes especially relevant:

Cross-synthesis with previous findings

L2.5 is by design an integrative session. Three tensions appear when connecting the physiological components of allostatic load and deserve to be made explicit:

Tension 1 — Metabolic buffer and cortisol phenotype may not correlate linearly

A woman can have a diverse diet and still show a collapsed cortisol phenotype because of accumulated historical allostatic load (ACEs, lactation, years of intense caregiving). Diet protects prospectively, but it does not necessarily reverse HPA setpoint recalibration quickly. Therefore: the model cannot assume that a better metabolic context automatically equals a better HPA state.

Tension 2 — Luteal load and cortisol phenotype are temporally offset

Luteal stress load updates per cycle (~28 days), whereas the diurnal cortisol phenotype is a relatively stable biotype whose inputs — AM/PM energy, type of insomnia — change daily. A month of high luteal load can coexist with a cortisol phenotype that does not move. The methodological consequence is that a composite phenotype should not be read as a single number: it is better to distinguish the current state (short window), the trajectory (deviation from each person's own baseline) and the cortisol phenotype as interpretive context. These are three elements on different time scales, not one.

Tension 3 — Ovarian context changes by hormonal stage

Markers linked to an active ovarian axis do not have the same interpretation in active cycling, anovulatory perimenopause or postmenopause. The same symptom may reflect different mechanisms depending on hormonal stage, so validation must stratify by cohort and avoid promising automatic transfer between groups.

Lua Labs hypotheses

Hypothesis 17 — Composite HPA phenotype as a longitudinal predictor of perimenopausal decompensation

Specific sub-statements:

• H17a — Dimensional structure. A composite HPA phenotype should factor-analyze into at least 3 interpretable latent factors: "biological buffer", "active allostatic load" and "circadian calibration". If a single factor emerges, the construct is redundant; if too many factors emerge, it is over-dispersed.
• H17b — Phenotypic validation. Diurnal cortisol phenotypes should predict different response patterns to dietary-behavioral and sleep interventions. If response is the same across phenotypes, the cortisol classification adds no useful signal.

Proposed mechanism:

Female HPA load integrates at least six arms: microbial buffer, vagal tone, cortisol phenotype, luteal stress, ovarian context and circadian calibration. Each arm operates on a different time scale:

• Microbial buffer: weeks-months scale, slow, may anticipate loss of HPA-axis buffering capacity before it manifests symptomatically.
• Vagal tone: weekly scale, autonomic mediator between stress, sleep and the HPA axis.
• Cortisol phenotype: relatively stable biotype that modulates how the system responds.
• Luteal load: per-cycle scale, the window of greatest vulnerability when progesterone and cortisol compete at GR signaling.
• Ovarian context: monthly or hormonal-stage scale, dependent on follicular activity and reproductive transition.
• Circadian calibration: daily scale, sensitive to sleep, light, food and caffeine.

The fact that dimensions operate on different time scales is a strength, not a weakness: it allows observation of signals that a single biomarker would not capture.

Confidence level: Medium.

How to validate:

With a complementary formal study (n=80, 90 days):

• Validate composite phenotype correlation vs diurnal cortisol slope.
• Validate composite phenotype vs nocturnal HRV (RMSSD, LF/HF).
• N=80 provides sufficient power (alpha=0.05, beta=0.20) to detect r=0.30 (small-medium effect).

With an external comparison dataset:

• MenoLife (Tariyal 2025) would be the natural comparison dataset if made available. In its absence: SWAN sub-cohort (public data) to validate the association classic AL ↔ retrospectively reconstructed longitudinal proxies.

Limitations:

1. Labeling "decompensation". Defining a clinically significant symptomatic episode from self-report data is debatable. A panel of endocrinologists + perimenopause gynecologists should co-define the criterion before analysis. Risk: if labeling has low inter-rater reliability, AUC will be artificially inflated or attenuated.

2. Construct validity. A composite phenotype may look robust by mixing many correlated signals. Validation must show that it adds information beyond sleep, subjective stress or isolated cortisol.

3. Transfer by hormonal stage. The hypothesis is stated for perimenopause. Extending it to active cycling, PCOS, POI or postmenopause requires validating each cohort separately.

Individual variability

A composite HPA-load phenotype must be interpreted in an individualized way. Similar signals can have different clinical meaning depending on:

• Hormonal stage: Sofía vs Carmen vs Rosa require different weighting (the active ovarian axis does not weigh the same in postmenopause).
• Cortisol phenotype: curves with high CAR, flat CAR or altered nocturnal nadir modify the reading of each component.
• Dietary acculturation: loss of Prevotella copri in urbanized Latinas (L1.4) can produce an apparently diverse diet with a functionally compromised microbial buffer.