Screening for pregnancy complications such as pre-eclampsia and gestational diabetes usually relies on one blood draw in the first trimester. Yet the metabolic changes that come before these diseases unfold continuously over weeks, and they differ from one woman to the next. Some pre-eclampsia starts early and is driven by the placenta. Some starts late and is driven by maternal factors. Nobody knows how much of the predictive and subtype-revealing information in a woman's metabolic trajectory a single draw captures, or how many draws, at which gestational weeks, would capture nearly all of it. We propose to answer this with a tool taken from antenna theory. Engineers designing "continuous" antenna surfaces showed that a smooth physical aperture carries only a finite number of independent information channels, set by how fast the eigenvalues of its propagation operator decay. They also showed that sampling the surface at a known spacing recovers almost all of that capacity. We treat the gestational metabolic trajectory the same way. The surface becomes the time axis, and the signal is whatever separates disease subtypes. From the eigenvalue decay we compute an information ceiling for screening schedules and a minimal set of draw times.
The result would be practical in two ways. It would tell clinical cohorts how to schedule draws. It would also tell modelers whether disease-progression models fail early in pregnancy because of their structure or simply because one draw contains too little information to distinguish subtypes.
Core Claim
The claim is that discriminative information about disease subtype in a continuous biomarker trajectory has a small effective number of degrees of freedom, and that this number can be estimated. If so, a schedule of roughly that many draws, placed at quadrature-like nodes derived from the operator's eigenfunctions, should reach nearly the full subtype-and-outcome discrimination of dense sampling. Uniform or convenience schedules with the same number of draws should fall short. In concrete terms, let $B(t,t')$ be the covariance across subtypes of the mean trajectory difference at gestational times $t,t'$, and let $W(t,t')$ be the within-woman covariance, which includes measurement noise and transient fluctuation. We solve the generalized eigenproblem
$$\int B(t,t')\,\phi_k(t')\,dt' = \mu_k \int W(t,t')\,\phi_k(t')\,dt',$$
where $\phi_k$ are discriminative temporal modes and $\mu_k$ are their signal-to-noise ratios. We define the effective degrees of freedom as $D_\epsilon = \#\{k : \mu_k > \epsilon\}$. For a draw schedule $S$, the ceiling on subtype discrimination is the sum of $\mu_k$ captured by $S$. The prediction is that this sum saturates once $|S| \approx D_\epsilon$ and the draws sit at the right nodes.
What Carries Over From Antennas, and What Does Not
Three continuous-aperture results supply the method. Electromagnetic Information Theory treats a channel as a compact linear operator between function spaces, so its information content is governed by eigenvalue decay. That is the abstraction we reuse. Performance Analysis for Near-Field MIMO: Discrete and Continuous Aperture Antennas derives the effective degrees of freedom of continuous and discrete apertures and shows that discrete spacing of about half a wavelength captures nearly all the multiplexing gain of the continuous surface. We want the analogous sampling-threshold statement for gestational time. Beamforming Optimization for CAPA-based Communications reduces an infinite-dimensional functional problem to a small correlation matrix using about 10 Gauss-Legendre quadrature nodes. This motivates our guess that a small number of well-placed nodes suffices.
The analogy breaks in one important place. In antennas the kernel follows from Maxwell's equations. In pregnancy it has to be estimated from data, from small cohorts, and it is non-stationary over gestation. So most of the scientific risk sits in estimating the kernel, not in the eigenanalysis.
Research Program
The model links trajectories to subtypes and stages. Each woman has a latent subtype $z$ and a continuous latent time shift $\tau$. The shift captures that the disease process starts at different gestational weeks. This is a continuous version of the stage variable in SuStaIn (missing-data SuStaIn), which infers discrete event orderings and handles missing modalities through a uniform likelihood. We drop the monotone event assumption because pregnancy biomarkers are non-monotone. For the within-woman structure we use the four-level design of the multi-resolution trajectory framework: a population mean, subtype B-spline curves, individual random effects, and a mean-reverting Ornstein-Uhlenbeck Gaussian process for transient fluctuation. This split matters because the eigenproblem needs $W$ to separate transient noise from subtype signal.
Fitting this model gives estimates of $B$ and $W$. From these we compute the modes $\phi_k$, their ratios $\mu_k$, $D_\epsilon$, and an optimized schedule. Uncertainty comes from bootstrapping over women, because with cohorts of tens to hundreds the eigenvalue tail will be unstable. We report $D_\epsilon$ only with intervals.
The biomarkers are those already shown to carry early signal. These are the leptin/ceramide (d18:1/25:0) ratio from the pre-eclampsia discovery study and the routine first-trimester panel from the Zhengzhou GDM nomogram: TG/HDL-C, APTT, AST, progesterone, and BMI. The routine panel matters because LC-MS/MS ceramides are expensive. A useful output is the ceiling for the cheap panel compared with the ceiling for the lipidomic ratio.
Baselines
The naive reference point is a single first-trimester draw fed to a static model, run exactly as in the Zhengzhou nomogram study: LASSO selection followed by logistic regression. That paper reports a validation AUC of 0.73 (95% CI 0.62–0.83), which falls from 0.85 in training. A second clinical reference is the sFlt-1/PlGF ratio. It scored AUC 0.52 in weeks 5–25 in the pre-eclampsia study, against 0.92 for the Lep/Cer ratio on 40 women. That 0.92 is the strongest reported early single-marker number, and any schedule must at least match the best single draw of that ratio on the same women.
The strongest existing trajectory method is the multi-resolution model. On 672 scleroderma patients it reports a 14.3% lower mean absolute error at a 4-year horizon and a 31% true-positive rate for detecting sharp declines, against 17% for baselines. It is the right competitor because it already separates transient from persistent change. It is fit on all available visits, while our schedules use a fixed budget of draws.
Missing-data SuStaIn is the strongest unsupervised subtype-and-stage competitor. The file gives no absolute accuracy figure for it, only "highest accuracy" up to 40% missingness.
The key ablation replaces our discriminative operator with ordinary functional principal components of the trajectories. The comparison uses the same number of draws, the same downstream classifier, and the same hyperparameter search budget. Functional principal components describe where trajectories vary, not where subtypes differ. No reported result was found on Lacuna for functional-PCA-based schedule design in gestational screening. A second ablation keeps the node count fixed and compares optimized nodes against uniform spacing.
Beating the baselines means two things. First, $k \approx D_\epsilon$ optimized draws must reach within a pre-registered margin of the dense-sampling AUC and subtype recovery. Second, they must exceed uniform and principal-component schedules of equal size, with the gain outside bootstrap intervals.
Why not just…?
Why not just use the best single biomarker at one draw? The Lep/Cer result shows that one draw can be strong. If our analysis finds $D_\epsilon \approx 1$, that is a real and publishable answer, because it would justify single-draw screening. The method is built to report this outcome, not to rule it out. What a single draw cannot do is separate early-onset from late-onset subtypes when they share an amplitude but differ in timing. The time-shift term tests exactly that case.
Why not just fit an XGBoost or logistic model on whatever visits exist? That answers a different question, namely how well a given dataset predicts. It cannot say what a different schedule would achieve. The Zhengzhou drop from 0.85 to 0.73 also shows that small static models overfit. Our ceiling is computed from fitted covariances and then checked by held-out subsampling, so it is evaluated rather than tuned.
Why not just use functional PCA plus standard Bayesian optimal design? This is the real threat. Our eigenproblem is close to functional discriminant analysis, and the antenna framing contributes a question and a sampling-threshold claim more than new mathematics. We do not argue around this. The principal-component ablation settles it directly. If variance-based schedules match discriminative ones, the discriminative operator adds nothing and we will say so. The remaining contribution would then be the empirical ceilings and the subtype-and-stage handling.
Proposed Experiments
The first check tests whether the ceiling is real. It runs on two datasets with dense longitudinal data: the Johns Hopkins scleroderma registry (4,992 lung-function measurements on 672 patients) and ADNI (789 participants). We estimate $D_\epsilon$ on training patients. We then subsample held-out patients' visits down to $k = 1, 2, \ldots$ draws, either at optimized nodes or uniformly, and measure subtype recovery and outcome AUC. The theory passes if the empirical saturation point falls within the bootstrap interval of $D_\epsilon$. It fails if performance keeps rising well past it.
The second check is pregnancy. We apply the same analysis to the 40-woman longitudinal pre-eclampsia cohort (weeks 5–29). Because the cohort is tiny, we treat these results as estimates with wide intervals, not as validation. We also apply it to any Zhengzhou cohort members with repeat draws. The question is whether timing-defined subtypes appear and whether $D_\epsilon$ is greater than one.
The third check is synthetic. We generate trajectories with known subtypes, time shifts, and eigenvalue spectra, at the sample sizes of the real cohorts. This measures how badly kernel estimation biases $D_\epsilon$ at n = 40 compared with n = 300, and so tells us how large a prospective cohort must be.
The decisive outcome comes from these checks together. Either a pre-registered pregnancy draw schedule of two to four draws closes most of the gap between one draw and dense sampling, or the ceiling shows that one well-chosen draw is already enough. Both outcomes would change how early-gestation cohorts are designed.
References
- Electromagnetic Information Theory: Fundamentals, Modeling, Applications, and Open Problems
- Performance Analysis for Near-Field MIMO: Discrete and Continuous Aperture Antennas
- Beamforming Optimization for Continuous Aperture Array (CAPA)-Based Communications
- Multimodal subtypes identified in Alzheimer’s Disease Neuroimaging Initiative participants by missing-data-enabled subtype and stage inference
- A Framework for Individualizing Predictions of Disease Trajectories by Exploiting Multi-Resolution Structure
- Early-pregnancy prediction of risk for pre-eclampsia using maternal blood leptin/ceramide ratio: discovery and confirmation
- Development and Validation of a Multidimensional Indicator-Based Risk Prediction Model for Gestational Diabetes Mellitus: A Nested Case-Control Study