Preeclampsia is a dangerous pregnancy disorder. Its root cause appears in the first trimester: the placenta fails to remodel the mother's uterine arteries into wide, low-resistance vessels. The mother's blood pressure, however, usually rises only after week 20. Low-dose aspirin helps only if it starts before about week 16. Prediction models built on routine electronic health records are good late in pregnancy and weak early. One such model reaches an area under the ROC curve (AUC) of 0.92 at week 37 but only 0.66 at week 4 (Mount Sinai trajectory study). Records cannot see the placenta.
One MRI technique can partly see it. Diffusion MRI measures how water molecules move inside tissue. When blood flows along organized vessels, or water moves along aligned fibers, the motion has a preferred direction. That directional bias is called anisotropy. In healthy pregnancies, placental blood flow is anisotropic in nearly every voxel, and tissue anisotropy concentrates at the uterine wall, exactly where artery remodeling happens (Placenta microstructure and microcirculation imaging with diffusion MRI). That study scanned only nine healthy women. The one study that scanned preeclamptic placentas threw away direction entirely (Combined diffusion‑relaxometry MRI).
We propose to test one falsifiable claim: among pregnancies that a records-based model already flags as high-risk, anisotropy at the uterine-wall boundary, measured at 14–16 weeks, changes risk estimates beyond what records, Doppler ultrasound and blood biomarkers already provide.
Core Claim
Most of the obvious combinations at this intersection either are already done or are the wrong bet. The obvious machine-learning version would bolt a "diffusion tensor specialist agent" onto the multi-agent language-model system MoMA (MoMA). MoMA turns images and lab panels into text summaries that a predictor model then reads. A skeptical reviewer would rightly point out two problems. Turning a handful of physical parameters into prose and back loses information. A plain regression on those same parameters would likely do as well. So we drop that design.
The real obstacle is a specific kind of data fragmentation. The informative modality, placental diffusion MRI, will only ever exist for a few hundred patients. The records exist for tens of thousands. The two populations barely overlap. Our claim has three parts:
- The imaging should be reduced to a few biophysically defined, uncertainty-weighted quantities at the uterine wall.
- These should enter as a correction on top of a records model trained on the large population, not as a jointly trained network.
- The imaging arm should be run as a second-stage test on records-flagged pregnancies. That makes the case count large enough to answer the question.
Three pieces are new. The first is measuring directional compartment parameters in at-risk pregnancies early enough for aspirin to matter. The second is ground-truthing what those parameters mean in thin, fibrous interfaces. The third is a fusion design sized honestly for a modality that lives in a small, separate population.
The other two directions play supporting roles rather than equal ones. Preeclampsia biology supplies the mechanistic anchor: failed artery remodeling and the trophoblast senescence program described by Nonn et al.. Trophoblasts are the placental cells that invade the uterine wall. Periodontal ligament stem cell (PDLSC) biology supplies a calibration tool. Multilayer sheets of periodontal-ligament-derived cells regenerate aligned collagen fibers anchored to bone and cementum (Iwata et al.). This is a thin, fibrous interface whose fiber alignment can be dialed up or down by steering the cells' fate (Fate decision of mesenchymal stem cells) and then verified by histology. Such a thin, graded, histology-checkable anisotropic layer is exactly what in vivo placental imaging lacks.
Research Program
Measurement. We use a multi-shell, multi-direction protocol modeled on the 2017 placental study. That study found, by Bayesian model comparison, that the "stick-zeppelin" model beats scalar and isotropic fits across all subjects. Stick-zeppelin is a two-compartment model: a fast, directional blood-flow part and a slower, possibly directional tissue part.
That study did not correct for maternal breathing motion, and its uterine-wall region is only a few millimeters thick. We therefore replace voxel-wise least-squares fitting with an amortized estimator: a network trained on simulated signals to output compartment parameters and their uncertainty in one pass. Training signals come from the forward model, corrupted with realistic noise and simulated respiratory displacement. This follows the motion-robust estimation and tensor-inpainting ideas in robust diffusion estimation and TW-BAG.
The per-patient imaging summary is kept small on purpose. It consists of the tissue-compartment anisotropy, the blood-fraction anisotropy, the perfusion fraction and principal-direction coherence. All are averaged over an automatically segmented uterine-wall band and weighted by the estimator's own uncertainty. Keeping it small echoes the diabetes-trajectory finding that a dozen strong features beat many noisy ones (INPC graph study).
Calibration. Before any clinical claim, we test whether the estimator recovers anisotropy and orientation in thin interfaces. We use two ex vivo substrates imaged on the same clinical scanner at matched resolution. The first is PDLSC-derived multilayer sheets grown under differentiation conditions that yield graded collagen alignment, with alignment quantified by polarized-light histology. The second is delivered placentas with histological scoring of artery remodeling. The PDL sheets give a controllable dose–response that delivered placentas cannot. The placentas give disease relevance that the sheets cannot.
Fusion under fragmentation. Let $f(x)$ be the log-odds of preeclampsia from a records trajectory model. We reimplement the Mount Sinai approach: functional principal components of vital-sign trajectories feeding gradient-boosted trees. This model is trained on the large cohort using records up to the scan week. Let $\theta$ be the imaging summary and $\sigma$ its uncertainty. The fused model is
$$\text{logit}\, p = f(x) + \beta_0 + \beta^\top \frac{\theta - \mu}{\sqrt{s^2 + \sigma^2}},$$
where $\mu$ and $s$ are the imaged cohort's mean and spread. The coefficient vector $\beta$ is the only thing fit on the small imaged cohort, with ridge regularization. With about four imaging parameters and roughly fifty events, this is a well-posed regression rather than an underpowered deep network. Noisy scans are automatically down-weighted because their uncertainty $\sigma$ enlarges the denominator.
Mechanistic link. At delivery, placentas from imaged pregnancies undergo histology of artery remodeling. They are also stained for the p300-linked senescence state and its secreted factors, Activin A and GDF15, following Nonn et al.. This tests whether in vivo uterine-wall anisotropy tracks the biology it is supposed to reflect.
Baselines
The naive reference point is guideline-based screening under the American College of Obstetricians and Gynecologists (ACOG) criteria. It reached an AUC of 0.66–0.67 in the Mount Sinai cohort (Mount Sinai trajectory study). We compute it on our imaged cohort from the same records.
The records-only trajectory model $f(x)$ is the direct comparison for the added value of imaging. Its reported early-pregnancy AUC is 0.66 at week 4. We retrain it on our site's records, frozen at the scan week, with the same tuning budget as all other arms.
The strongest competitor is the nuMoM2b-style combination of records, uterine artery Doppler pulsatility index, and the circulating proteins PlGF and soluble endoglin (PEPrML, nuMoM2b). It reports an AUROC of 0.72–0.77 across visits, with biomarkers adding 7% over clinical factors. We add Activin A and GDF15, since Nonn et al. report they improve on Doppler alone. We give this arm the identical offset structure on $f(x)$ with the same number of free coefficients, the same cross-validation folds and the same regularization search. The only thing that differs from the imaging arm is what $\theta$ contains.
The ablation that isolates our key ingredient refits the same acquisition with isotropic models: scalar diffusion coefficient and isotropic two-compartment fits. It uses the same estimator, region and fusion. If directional information matters, this arm should lose.
We also run a direct-fitting arm that uses conventional least-squares parameter fits instead of the amortized estimator, to isolate motion robustness. A MoMA-style text-summary arm is included as a representative of the agent approach to fragmented data.
No reported result was found on Lacuna for preeclampsia prediction AUROC from placental diffusion MRI of any kind. The imaging arms therefore have no published number to beat. The bar is the biomarker arm on the same patients.
"Beating" it means one of two things. The first is a gain in AUROC over records-plus-Doppler-plus-biomarkers whose 95% bootstrap interval excludes zero. The second is a net reclassification improvement large enough to change aspirin decisions for at least one in ten flagged patients. Either must hold on the held-out site, without a worsening of the predictive equality ratio across racial groups. The PEPrML baseline for that ratio was 1.31, reduced to 1.14 after threshold adjustment.
Why not just…?
Why not just use Doppler plus blood biomarkers, which are cheap and already validated? This is the real threat, and the experiment is built to let it win. Doppler measures resistance in one large artery. Biomarkers measure what the placenta releases into the blood. Neither localizes which part of the uterine wall failed to remodel. Nonn et al. still needed senescence proteins to rescue Doppler's early underestimation. If the imaging term adds nothing beyond this arm, we will report that placental diffusion MRI is not worth its cost for early screening. That would be a useful negative result, given how often placental MRI is proposed.
Why not just use scalar diffusion or the diffusion-relaxometry method, which every scanner supports? Scalar diffusion and the combined diffusion-relaxometry (ZEBRA) approach did separate some preeclamptic placentas, but in late pregnancy and in a cohort of only 17. Its authors named the missing directional information as a limitation. Bayesian model comparison in the 2017 study rejected isotropic fits everywhere. Model fit, however, is not prediction. The isotropic ablation, run on the same acquisition, settles whether direction carries disease signal or merely fits the data better.
Why not just add a diffusion specialist agent to MoMA? MoMA's specialists were validated only on chest X-rays and lab panels. The system reached macro-F1 near 0.85 on trauma stratification (MoMA). Its value lies in avoiding paired training data for high-dimensional inputs. Our imaging input is four physical numbers, so there is nothing for the agent to compress, only room to hallucinate. We run the agent arm anyway. We expect it to match the offset model at best.
Why not just train an end-to-end network on the raw parameter maps? With roughly fifty events, a convolutional network on 3D maps would overfit. The pruning result from the diabetes study and MoMA's own argument against joint training both point the same way.
Why not just calibrate on delivered placentas and skip the PDL sheets? Delivered placentas are the more relevant substrate, and we use them. They offer only whatever alignment disease happened to produce. PDL sheets add a controlled, graded series of alignments in a layer of similar thickness. If the estimator's accuracy on placentas alone proves sufficient, the sheets become an optional check rather than a requirement.
Proposed Experiments
Calibration. In PDL sheets with histology-graded alignment, estimated tissue anisotropy should rise monotonically with measured collagen alignment. We assess this by rank correlation. Orientation error at the clinical scanner's resolution should fall below the angular spread seen in vivo at the uterine wall. The amortized estimator should beat least-squares fitting on simulated signals with injected breathing motion. If these fail, the clinical arm is uninterpretable and we stop.
Prospective cohort. Nulliparous pregnancies in the top risk decile of the records model at 12 weeks are invited for MRI at 14–16 weeks and again at 20–24 weeks. The target is about 300 women, split across two sites so one can serve as external validation, mirroring the Mount Sinai MSH-to-MSW design. Enrichment should raise preeclampsia prevalence well above the population rate and yield enough events to fit the fusion coefficients. Early-pregnancy placental MRI is less established than the 27–38 week scans in the 2017 study. Scan feasibility and anisotropy detectability at 14 weeks are therefore measured as a first-stage outcome.
Decisive checks. The primary test is the gain in AUROC and reclassification of the anisotropic imaging term over the strongest biomarker arm at 14–16 weeks, on the held-out site. The secondary test is the same comparison against the isotropic ablation.
The mechanistic test asks whether uterine-wall anisotropy in vivo correlates with histological remodeling and with the p300/senescence markers in the matching delivered placenta. A predictive gain without that correlation would suggest the signal comes from something other than remodeling failure. A correlation without a predictive gain would suggest the measurement is sound but redundant with the blood biomarkers.
Finally, fairness is audited with PEPrML's predictive equality ratio, so that an imaging arm that helps on average does not quietly widen the error gap between groups.
References
- Improving preeclampsia risk prediction by modeling pregnancy trajectories from routinely collected electronic medical record data
- Placenta microstructure and microcirculation imaging with diffusion MRI
- Combined diffusion‐relaxometry MRI to identify dysfunction in the human placenta
- MoMA: a mixture-of-multimodal-agents architecture for enhancing clinical prediction modelling
- Disturbed trophoblast transition links preeclampsia progression from placenta to the maternal syndrome
- Periodontal regeneration with multi-layered periodontal ligament-derived cell sheets in a canine model
- Fate decision of mesenchymal stem cells: adipocytes or osteoblasts?
- Robust estimator framework in diffusion tensor imaging
- TW-BAG: Tensor-wise Brain-aware Gate Network for Inpainting Disrupted Diffusion Tensor Imaging
- Predicting onset of complications from diabetes: a graph based approach
- A comprehensive and bias-free machine learning approach for risk prediction of preeclampsia with severe features in a nulliparous study cohort