A Target-Separated Recoverability Audit for Frozen Foundation-Model Embeddings, Demonstrated on AMH in IVF
This repository audits where AMH and ovarian-reserve information is recoverable from frozen foundation-model embeddings across public IVF-related data, including DINOv2, DINOv3, CLIP, USF-MAE, and related encoders. The analyses use predictive V-information probing, conditional age probing, control tasks, MDL summaries, and model-class sufficiency checks with TOST. The contribution is the target-separated audit protocol and the reusable reserve_audit package, not a new model.
Target-separated recoverability atlas. Held-out fixed-linear recoverability R/H(Y), grouped by registered target kind. Rows are never ranked across target kinds; the ovary-US phenotype node is a phenotype positive control, not an ovary-to-serum-AMH test.
1. AMH-tertile recoverability depends on modality and survives age-conditioning only for endocrine trajectories. In Brigham/Leahy endocrine trajectories, recoverability changed from marginal R/H=0.100225 to age-conditioned R/H=0.087189 (2000-shuffle conditional permutation p=0.0004997501249375312). In Wang embryo morphology it changed from marginal R/H=0.004202 to age-conditioned R/H=-0.000044 (p=0.597015).
Marginal vs age-conditioned R/H, and the Brigham conditional statistic against its 2000-shuffle age-permuted null.
2. The same frozen-encoder pipeline recovers an ovary-US phenotype positive control, so the embryo result is a measured boundary rather than a failed pipeline check: DINOv2 R/H=0.3401 and USF-MAE R/H=0.2974 on FUID, while ten embryo encoders stay near zero for serum-AMH tertile.
Frozen-encoder benchmark: embryo-to-serum-AMH is null across encoders; ovary-US phenotype is recovered by the same V-information probe.
3. Clinical concordance points in the same direction. A 6-cohort AMH meta-analysis found a quantity-marker association of r=0.4054 (95% HKSJ CI 0.2833 to 0.5145, p=0.0005032429233777444); the quality-marker pool was near zero at r=0.0156 (95% HKSJ CI -0.0262 to 0.0575, p=0.35808119845986486). Quantity effects were positive in 6/6 cohorts; quality effects were near zero in 5/5 estimable cohorts.
AMH behaves as a quantity marker, not a per-embryo-quality marker, across IVF cohorts.
These are modest aggregate effects. The age-conditioning results are single-cohort and adjust for age only. They should be read as a representation-analysis audit, not as an image-to-serum-AMH predictor.
reserve_audit/ reusable audit package (registry, recoverability, sufficiency, CLI)
scripts/ benchmark, hardening, verification, and figure generators
results/diagnostics/ aggregate diagnostic JSON files used for the public claims
figures/ public figures generated from aggregate diagnostics
requirements.txt minimal Python dependencies
Install the minimal dependencies:
pip install -r requirements.txtThe diagnostics JSON files in results/diagnostics/ back the numeric claims above. The encoder-benchmark figure regenerates from the aggregate benchmark JSON, and the benchmark values can be re-checked:
python scripts/make_fig_encoder_benchmark.py
python scripts/verify_benchmark.pyRaw cohort data and cached embeddings are not redistributed here; users should obtain those assets from their original sources and licenses. This release contains the protocol code, aggregate diagnostics, figures, and generators.
Public cohorts used by the audit include Brigham/Leahy 2021, Mendeley 5k IVF, Wang embryo-image data, FUID ovary-US, Clomiphene DOR, PGT-A euploidy DOR, Embryoscope hr-NGS KIDScore, and Heavy-metal DOR. Users should obtain each cohort from its original source.
This is representation-analysis and probing methodology demonstrated on AMH and IVF. Image-to-serum-AMH recovery is future work gated on paired ovary-US plus continuous-AMH data, and is not claimed here. A master's thesis using this audit is in preparation.
MIT (see LICENSE).