Toward a Dynamic Ovarian Recruitment Phenotype in Assisted Reproduction

1. The Clinical Problem Is Not a Lack of Serial Measurements

Ovarian response assessment already combines information obtained before and during stimulation. AMH and AFC are useful predictors of low and excessive response and contribute to individualization of treatment, but they do not directly measure oocyte competence or the chance of a live birth (Hendriks et al., 2005; Broer et al., 2009; Broer et al., 2011; Dewailly et al., 2014; Iliodromiti et al., 2015; Liu et al., 2023). The updated ESHRE guideline recommends either AMH or AFC for predicting response; this established clinical role should not be displaced by an unvalidated imaging construct (ESHRE Guideline Group on Ovarian Stimulation, 2026). Individualized follitropin delta dosing studies also show the clinical importance of established reserve-guided treatment strategies (Nyboe Andersen et al., 2017; Ishihara and Arce, 2021). AFC measurements have recognized acquisition and observer variability, and AMH has its own biological and assay characteristics (Broekmans et al., 2010; Moolhuijsen and Visser, 2020).

During a stimulation cycle, clinicians also review repeat follicle counts, size distributions, hormonal measurements and administered gonadotropin doses. These observations are genuinely longitudinal. The narrower unresolved question is whether retaining the identity of each follicle across observations adds information that those existing visit-level measurements cannot convey. A better prediction based on a later scan is not, by itself, evidence for a new phenotype; the new information must survive comparison at the same time point and using the same scanning history.

The distinction matters when a patient with apparently adequate baseline markers has an unexpectedly low oocyte yield. Baseline reserve and realized response are related but not identical, and ovarian response can differ between individuals and between cycles of the same individual (van Disseldorp et al., 2010; Rombauts et al., 2015; Arvis et al., 2022; Hochberg et al., 2026). Indices such as the Follicular Output RaTe (FORT), follicle-to-oocyte index (FOI) and ovarian sensitivity index (OSI) quantify aspects of realized response, but many of their inputs are only known after much of stimulation or retrieval has occurred (Genro et al., 2011; Huber et al., 2013; Alviggi et al., 2018). They are useful comparators for completed cycles, not automatically fair predictors at an early landmark.

2. Central Hypothesis: Identity Is Information, Not Merely Annotation

We define the dynamic ovarian recruitment phenotype as the joint temporal behavior of an identified, ultrasound-visible antral follicle cohort during a stated observation window and under a stated stimulation regimen. In this definition, 'recruitment' means entry into, and progression within, the measurable antral cohort. It does not describe activation of primordial follicles, which cannot be followed individually by routine transvaginal ultrasound in living women. 'Phenotype' denotes a candidate measurable pattern, not an established patient classification or an intrinsically fixed ovarian trait.

The primary hypothesis (H1) is that preserving follicle identity reveals predictive information that is irretrievably discarded when each visit is reduced to unordered counts and size distributions. Let B represent baseline clinical characteristics, V(t) all routine visit-level observations available through an early landmark t, G(t) recorded treatment exposure and dose decisions through t, I(t) the identity-linked trajectories derived from the same scans, and Y the number of mature oocytes. The operational null hypothesis is that adding I(t) to B, V(t) and G(t) yields no clinically meaningful improvement in out-of-sample prediction of Y. Rejection of this null requires a prespecified improvement under external validation, not simply a statistically significant association or a better fit in the development set.

The reason identity might matter is elementary but testable. Imagine two visits at which the observed follicle diameters are 5 and 9 mm, followed by 7 and 11 mm. Visit-level distributions are identical whether the follicles grow by 2 mm each or whether one changes by 6 mm and the other by -2 mm. The latter assignment could reflect genuine heterogeneity or an erroneous match; linking follicles by size rank alone would always select the first assignment and could never register the second. Only reliable identity linkage can distinguish these possibilities; whether that distinction predicts clinical outcomes is precisely the proposition to test. This numerical example is illustrative, not observed patient data.

A second hypothesis (H2) is that some trajectory-derived information has a reproducible patient component after conditioning on stimulation protocol, medication exposure, time since stimulation start, age and measurement uncertainty. H2 is stronger than H1: a within-cycle predictor may be useful without defining a stable individual phenotype. A third, exploratory hypothesis (H3) asks whether registered antral-follicle location or local neighborhood adds information beyond temporal trajectories. Neither H2 nor H3 follows from H1, and they should be evaluated separately.

3. Physiological Motivation, Without Overstating the Mechanism

Human follicle development includes early processes regulated largely by intraovarian signals and later stages with increasing gonadotropin responsiveness (Gougeon, 1996; McGee and Hsueh, 2000; Gougeon, 2010). Repeated ultrasound studies show more than one antral follicular wave during many natural menstrual cycles and reveal heterogeneous growth and regression patterns (Baerwald et al., 2003; Baerwald et al., 2009; Baerwald et al., 2012; Baerwald and Pierson, 2020). These observations make a temporally structured response plausible. They do not establish that the structure is clinically useful during controlled stimulation, which changes the endocrine environment and the observed cohort.

The reserve underlying reproductive ageing is more complex than the visible antral compartment. Human histological work documents age-related decline and substantial variability in non-growing follicles (Charleston et al., 2007; Hansen et al., 2008; Wallace and Kelsey, 2010). Molecular work implicates local regulatory networks, including PI3K-AKT-FOXO3 signaling, in primordial activation (Adhikari and Liu, 2009). Such mechanisms cannot be inferred from trajectories of millimeter-scale antral follicles. This distinction protects the proposal from confusing a clinically observable response pattern with microscopic ovarian biology.

D’Angelo et al. (2026) used whole-ovary imaging and modelling to describe age-associated oocyte distributions, local density relationships and inferred activation dynamics in mice. The study is a compelling source of questions about spatial organization, but its mouse oocyte-size distributions, including the secondary peak near 60 micrometers, cannot be treated as a threshold for human antral follicles or as evidence that ultrasound reveals primordial activation. The reported human ovarian cortex analysis supports technical feasibility of tissue imaging, not longitudinal human activation kinetics. In our framework, spatial hypotheses remain downstream, exploratory correlates unless independently validated in women.

4. Translating the Hypothesis into Measurable Features

A measurement framework should begin with a small, interpretable set of identity-dependent features rather than a high-dimensional imaging signature. Candidate features are: (i) persistence of baseline-detected follicles with credible continued growth by landmark t; (ii) distributions of individual growth velocity normalized by actual time between scans; and (iii) cohort synchrony, such as dispersion of within-follicle slopes or the stability of growth ranks. Their definitions and minimum measurable follicle size must be fixed before outcome analysis, with sensitivity checks across plausible detection thresholds.

These features must be demonstrably identity dependent. For example, a difference in average follicle size or total count between day 3 and day 6 can already be obtained without tracking. A candidate trajectory feature should therefore be evaluated after providing the comparator with all visit-level counts, quantiles, size histograms, time intervals and relevant clinical variables. As a negative control, follicle identities can be permuted between adjacent visits while preserving the visit-level distributions; any claimed trajectory benefit that persists under such scrambling requires scrutiny. Permutation should be performed within a prespecified scheme that respects measurement constraints, and should not be used as a surrogate for a genuine independent validation set.

Technical feasibility is not the missing claim. A 2021 proof-of-concept using 3D transvaginal ultrasound described automated tracking across IVF visits, including follicles above approximately 4 mm, but did not establish clinical incremental prediction (Srivastava et al., 2021). Nor is the use of serial follicle data new. Machine-learning models trained on routine monitoring scans predict trigger timing and over-response (Robertson et al., 2021), an interpretable model determines the optimal day of trigger from follicle-size profiles (Fanton et al., 2022), and an explainable model developed across eleven clinics identified which trigger-day follicle sizes contribute most to mature-oocyte yield (Hanassab et al., 2025). Each of these treats every scan as an unordered profile of sizes; none preserves which follicle is which. The closest attempt to model individual follicles, an agent-based analysis of more than 430,000 follicles, assigned identity by rank-matching sizes between consecutive scans under the explicit assumption that the largest follicle remains the largest (Hramyka et al., 2026). Its endpoint was the follicle-size profile at the end of stimulation rather than oocyte yield, and rank-matching is precisely the identity-free surrogate that H1 proposes discards information. Reviews of AI in follicle monitoring further caution that serial inputs, truly time-aware analysis and benefit for an actual decision are distinct levels of evidence (Kuang et al., 2026; Liu et al., 2026). The proposed construct is deliberately algorithm-agnostic: reliable tracking may use human review, image registration or machine learning; added clinical value must be shown independently.

Table 1. Candidate identity-linked features and the test they must pass.

Candidate component	Predeclared measurement	Falsification / negative control
Identity-linked persistence	Fraction of reliably matched baseline follicles with growth beyond a predeclared measurement-noise band by landmark t.	Compare with baseline cohort count and contemporaneous size distributions; stress-test detection thresholds.
Individual growth velocity	Per-follicle size change divided by actual elapsed time; robust summaries of within-cohort heterogeneity.	Control for visit timing and size histograms; exclude uncertain tracks in sensitivity tests.
Cohort synchrony	Dispersion and rank persistence of matched follicle trajectories.	Show incremental value beyond visit-level distribution changes and identity-permutation controls.
Spatial organization (exploratory)	Registered antral position or neighborhood only when repeatable across scans.	Require registration reliability and added value beyond temporal trajectories; no claim about primordial activation.

Table 1. Candidate identity-linked features and the test they must pass.

Candidate component	Predeclared measurement	Falsification / negative control
Identity-linked persistence	Fraction of reliably matched baseline follicles with growth beyond a predeclared measurement-noise band by landmark t.	Compare with baseline cohort count and contemporaneous size distributions; stress-test detection thresholds.
Individual growth velocity	Per-follicle size change divided by actual elapsed time; robust summaries of within-cohort heterogeneity.	Control for visit timing and size histograms; exclude uncertain tracks in sensitivity tests.
Cohort synchrony	Dispersion and rank persistence of matched follicle trajectories.	Show incremental value beyond visit-level distribution changes and identity-permutation controls.
Spatial organization (exploratory)	Registered antral position or neighborhood only when repeatable across scans.	Require registration reliability and added value beyond temporal trajectories; no claim about primordial activation.

5. A Prospective Test Designed to Be Difficult to Pass

The primary study population should be a prospectively recruited, multicentre stimulation cohort with the intended-use population and reserve-marker strata declared in advance. An illustrative first application is identifying a lower-than-anticipated mature-oocyte yield in women whose baseline AMH and AFC suggest adequate response. 'Adequate' and 'unexpectedly low' must be operationalized a priori using age, assay and clinical context; thresholds cannot be chosen retrospectively to maximize model separation. Across the cohort, mature-oocyte count is a natural primary continuous outcome. A low-yield event can be a prespecified secondary endpoint, not the sole outcome used to construct the phenotype.

A practical acquisition plan would obtain baseline and scheduled stimulation 3D transvaginal ultrasound volumes, while recording routine two-dimensional measurements, hormonal tests, prescribed and administered gonadotropins, dose changes and scan timing. For illustration only, a prediction landmark might be stimulation day 6 with baseline, day-3 and day-6 information collected in prespecified windows. The choice of day and visit frequency requires feasibility evaluation and is not a recommendation to change routine care. The same acquired 3D examinations should generate both the identity-free and identity-linked summaries in the primary comparison, so any gain cannot simply be attributed to one model having a better modality or more scans. A separate pragmatic comparison against real-world routine monitoring would establish the deployment value.

At the common landmark, three prespecified models should be tested, together with one intermediate variant. Model A includes baseline predictors such as age, AMH, AFC and planned regimen. Model B adds the full routine monitoring history available by the landmark: follicle counts and size distributions at each visit, scan timestamps, relevant hormone results and treatment exposure to date. Model B′ adds to Model B a rank-matched pseudo-identity, in which follicles are linked between consecutive scans by size order alone, as in Hramyka et al. (2026); it requires no tracking technology and can be computed from routine two-dimensional measurements. Model C receives precisely Model B's information plus reliable identity-linked trajectory features derived from the same scans. Model C versus B is the decisive experiment; C versus A alone is not sufficient. Model C versus B′ is the economic experiment: if genuine identity linkage does not outperform rank-matching, the costly component of the proposal, three-dimensional acquisition and tracking, is unnecessary, and that would itself be a useful result. Model flexibility, hyperparameter search, handling of missingness and validation effort should be comparable, and all optimization must occur within training folds. Neither model may use later scans, trigger-day information, total cycle dose accrued after the landmark or oocyte-retrieval details as predictors.

The treatment regime cannot be treated as an incidental covariate. Gonadotropin dose and formulation influence future growth, yet clinicians modify dosing partly in response to prior scans and hormones. This feedback means the observed pattern is a treatment-conditioned state, and simple adjustment may not establish causality. At minimum, exposure and dose decisions must be time stamped and included consistently in B and C; analyses should assess variation by protocol and changes in dose. Claims about intrinsic responsiveness or hypothetical outcomes under alternative dosing would require appropriate causal designs that are beyond this proposed first validation (Liu et al., 2026).

Validation should be separated by patient, not by scan or cycle, to prevent the same woman's data from entering both training and test sets. The study should specify a priori the primary error measure for mature-oocyte count, confidence intervals and an acceptable magnitude of improvement; secondary evaluations should include calibration and error distribution, and discrimination and calibration for any categorical low-yield outcome. Centre-level and equipment-level external testing, subgroup analyses and monitoring of failure-to-track rates are essential. Sample-size planning should reflect the intended effect, number of centres, outcome distribution, repeated cycles and expected cancellation rate rather than a convenient number of ultrasound examinations.

Cycles without retrieval create a consequential missing-outcome problem, not a reason for routine exclusion. Cancellation, conversion or failed retrieval should be reported, and the estimand for mature-oocyte count among completed retrievals distinguished from outcomes defined per initiated cycle. The main prediction model must not silently exclude the most difficult patients. Missing scans, partial ovaries and uncertain exposure histories should be handled under prespecified rules; sensitivity analyses should test whether the result depends on complete-case selection.

6. Measurement Validity and Reproducibility Are Separate Experiments

Before assigning biological meaning, identity labels must be validated. Repeated acquisitions in a subset, independent manual adjudication, inter-observer and inter-device comparisons, and measurements of follicle-detection and matching error can establish whether trajectories are trustworthy. A follicle first appearing above the ultrasound detection threshold may have been present but invisible at an earlier visit. An apparent contraction or disappearance may instead be a segmentation or matching failure. Uncertain identities should be flagged or probabilistically represented, never filled with confident trajectories by default. Acceleration and changing spatial neighborhoods should remain exploratory unless sufficiently dense scans support their estimation.

Alongside the identity-permutation control described in Section 4, sensitivity analyses should vary plausible segmentation and matching errors, exclude uncertain tracks, and repeat the comparison using human-reviewed identities. If a model's performance depends heavily on imperfect matching, imaging device or scan intervals, it is evidence of a measurement artifact or restricted generalizability rather than a robust ovarian phenotype.

To evaluate H2, a separate repeated-cycle cohort should distinguish within-patient from between-patient variation under comparable protocols, while modelling age, time, treatment exposure and cycle-level effects. Retaining a stable component does not imply perfect cycle-to-cycle agreement; indeed, important inter-cycle differences have been documented (Rombauts et al., 2015; Hochberg et al., 2026). Patients who repeat treatment may be selected on prior outcomes, so selection into repeat cycles requires attention. If prediction improves within a cycle but between-cycle stability is negligible, the correct description may be a dynamic monitoring signature rather than a patient-specific recruitment phenotype.

7. What Would Support, Weaken or Refute the Proposal?

The hypothesis should not be protected by continually adding features, moving the landmark or changing outcomes until a positive finding appears. A successful primary study would show that a small prespecified identity-linked feature set outperforms both Model B and Model B′ in a genuinely external cohort, with acceptable calibration and tracking reliability. This would support H1 as a predictive statement; it would not by itself establish patient-level stability, underlying follicle-activation biology or a better clinical outcome.

Several results would count against H1: no meaningful incremental gain over the full monitoring history; an advantage explainable solely by better imaging, later visits, flexible modelling or post-landmark leakage; loss of performance under independently verified identity labels; or dependence on one centre or a subset of easily tracked follicles. H2 would be weakened if repeated-cycle features lacked reproducible patient-level variation under comparable treatment conditions. H3 would be weakened if spatial associations vanished after registration quality control or adjustment for cohort size, anatomy and treatment. Each of these would be a valuable negative finding, identifying where added monitoring complexity is unnecessary.

The scientific claim should be distinguished from clinical utility. An early warning might prompt counselling or clinician review, but there is no evidence from this hypothesis alone that changing dose, trigger timing or cancellation decisions improves outcomes. Response counts are intermediate endpoints; improved prediction of mature-oocyte yield cannot be presumed to improve embryo competence or live birth (Broer et al., 2009; Fleming et al., 2015; Sunkara et al., 2011). Established approaches to poor and low-prognosis response, including Bologna and POSEIDON, also address dimensions not replaced by tracking (Ferraretti et al., 2011; Conforti et al., 2019). Only after predictive validation should a defined decision rule be tested prospectively for net benefit, safety, clinician workload, patient burden and cost.

Table 2. Distinct claims and outcomes that would weaken them.

Claim	Decisive test	Result that would weaken the claim
H1: incremental prediction	Externally validated Model C versus Model B at the same landmark, with the same scans and treatment information.	No prespecified clinically relevant gain; gains disappear under matching or leakage controls.
H1′: identity beyond rank-matching	Model C versus Model B′ (rank-matched pseudo-identity computed from the same scans).	No gain over rank-matching; three-dimensional tracking adds cost without information.
H2: patient component	Repeated-cycle variance analysis, conditional on protocol, exposure and time.	No reproducible between-patient component, even if within-cycle prediction is useful.
H3: spatial extension	Reproducible registration and incremental signal beyond temporal features.	Signal disappears after location-error correction or external testing.
Clinical utility (later stage)	Prospectively evaluated decision rule with net-benefit and safety outcomes.	Prediction changes risk scores but does not improve a defined clinical decision.

Table 2. Distinct claims and outcomes that would weaken them.

Claim	Decisive test	Result that would weaken the claim
H1: incremental prediction	Externally validated Model C versus Model B at the same landmark, with the same scans and treatment information.	No prespecified clinically relevant gain; gains disappear under matching or leakage controls.
H1′: identity beyond rank-matching	Model C versus Model B′ (rank-matched pseudo-identity computed from the same scans).	No gain over rank-matching; three-dimensional tracking adds cost without information.
H2: patient component	Repeated-cycle variance analysis, conditional on protocol, exposure and time.	No reproducible between-patient component, even if within-cycle prediction is useful.
H3: spatial extension	Reproducible registration and incremental signal beyond temporal features.	Signal disappears after location-error correction or external testing.
Clinical utility (later stage)	Prospectively evaluated decision rule with net-benefit and safety outcomes.	Prediction changes risk scores but does not improve a defined clinical decision.

8. An Evidence Ladder, Not a Clinical Recommendation

The translational sequence is therefore: reliable identity linkage first; then incremental information at a prespecified landmark against both identity-free and rank-matched comparators; then a repeatable patient component; and only then a decision rule in a prospective impact study. Spatial analysis may proceed as exploratory physiology but must not be used to rescue a negative primary result. Bilateral comparison may be a useful within-patient probe, given reports of possible left-right differences in response despite similar baseline indicators, but the mechanism is unresolved and replication is needed (Abrahami et al., 2020).

This approach is intentionally compatible with either a simple transparent statistical model or a more complex learning method. The biological hypothesis does not depend on a particular neural network, proprietary scanner or 3D reconstruction algorithm. If preserving identity is useful, its value should be demonstrable with proper controls. If ordinary monitoring suffices, the same controls should allow the proposal to fail decisively. In both cases, a fair test advances understanding of what information is actually required for ovarian-response assessment.