Clinicians rarely ask whether one number has changed. They ask whether a patient's overall pattern of change between visits is unusual for someone like them. In Alzheimer's disease, that pattern covers hippocampal shrinkage, ventricular enlargement, entorhinal thinning and cognitive decline all at once. The best calibrated tools we have answer a narrower question. They check one biomarker at a time and flag a patient when that biomarker's trajectory moves outside a band that is guaranteed to hold 90% of normal trajectories. This proposal replaces "is this biomarker's slope extreme?" with "how extreme is this patient's whole vector of interval change, compared with normal agers who had the same starting point and the same gap between visits?" The answer comes with an error guarantee that does not depend on any distributional assumption.
The tool is the multivariate rank built from optimal transport. Optimal transport is a way of warping one probability distribution into another at minimal cost. Here it warps the cloud of normal change vectors onto a uniform ball. After the warp, a patient's distance from the center of the ball is a rank between 0 and 1: 0 is typical change, and values near 1 are extreme. The direction from the center is a "sign" that says which combination of biomarkers is driving the deviation. Thresholding the rank gives a progression alarm with a controlled false-alarm rate. The sign gives a readable direction of progression, which a single score cannot give.
Core Claim
Conditional center-outward ranks of multi-biomarker change vectors can do something current tools cannot. The ranks are learned as a transport map conditioned on baseline state and time since the last visit, and then calibrated by conformal prediction, a finite-sample recalibration step. At the same finite-sample false-alarm rate on cognitively normal subjects, they should detect conversion from mild cognitive impairment (MCI) to Alzheimer's disease more often and earlier than univariate conformal bands. They should also outperform the two natural multivariate shortcuts: a conformalized Mahalanobis ellipsoid, and a max-over-biomarkers normalized residual.
The claim is falsifiable and deliberately narrow. Any gain over the ellipsoid must come from normal change being non-elliptical once we condition on baseline and time gap. Examples are curved or skewed shapes, or a mixture by APOE4 status. The program tests that premise directly rather than assuming it.
What Exists and Where This Sits
The mathematical machinery comes from the optimal-transport ranking literature. Monge–Kantorovich depth, quantiles, ranks and signs defines center-outward ranks and signs by mapping data to a uniform ball. Vector quantile regression: An optimal transport approach extends this to conditional settings. It writes the map as the gradient of a convex potential, a function whose gradient generalizes "monotone increasing" to many dimensions, and this guarantees that quantile contours never cross. We reuse both results unchanged.
Three recent conformal methods make these ranks practical. Neural Optimal Transport Meets Multivariate Conformal Prediction parameterizes the potential with input-convex neural networks, which are networks constrained to be convex in their input. It amortizes fitting across conditioning inputs, and it "pulls back" ball contours into curved prediction regions that are smaller than ellipsoids on bio and blog at 90% coverage. Optimal transport-based conformal prediction uses discrete transport ranks, plus a k-nearest-neighbor local variant (OT-CP+). It notes the cubic cost of that approach and reports poor worst-set coverage for ellipsoids. Geometric Conformal Prediction with Spatial Ranks and Multivariate Quantiles substitutes cheaper and more outlier-robust spatial ranks, at the price of affine equivariance. All three evaluate on static tabular or image benchmarks. None conditions on elapsed time between visits, and none uses the rank as a test statistic for change within a patient. Those two elements are what we add.
The motivating clinical work is Uncertainty-Calibrated Prediction of Randomly-Timed Biomarker Trajectories with Conformal Bands. It builds simultaneous conformal bands for irregularly timed trajectories on a harmonized ADNI and BLSA cohort of 2,200 subjects and reaches nominal 90% coverage. It defines a Rate of Change Bound (the slope of the lower band), which identifies 17.5% more MCI progressors than point-prediction slopes. Its authors state that it handles one biomarker at a time and that joint guarantees across coupled biomarkers are future work. Our proposal is that future work, with one change of framing. We score whole change vectors, not whole trajectories, because the clinical decision is made at each new visit.
Two other works constrain the design. Unscrambling disease progression at scale is the only paper in the file that uses optimal transport for disease progression. It uses entropic transport to relax a permutation over event orderings from cross-sectional data. It builds no rank maps and leaves calibrated uncertainty open. We treat it as a complementary population-level ordering, not as a competitor for patient-level alarms.
The longitudinal imaging models are relevant in a different way. DeltaNet, BioViL-T and CoCa-CXR argue that the between-visit change, not the current snapshot, carries the diagnostic signal. That supports our choice to rank change vectors. These models are, however, deterministic classifiers with no coverage guarantees. We use their "model the change" principle but not their architectures. We also keep chest radiographs out of the main study, because MS-CXR-T labels are three-way categorical and are not continuous change vectors.
Research Program
Each subject contributes pairs of consecutive visits. For each pair, $x$ is the baseline state (the regional volumes at visit one, age, sex and APOE4 status) and $g$ is the time gap in years. The change vector $y \in \mathbb{R}^d$ collects annualized changes in a small, pre-registered set of biomarkers. These are hippocampal volume, ventricular volume and entorhinal volume from the 145 harmonized regions, plus a cognitive score where ADNI provides one, so $d$ is 3 to 5.
On cognitively normal subjects who stay normal, we fit a conditional convex potential $\psi(\cdot\,; x, g)$. Its gradient pushes the uniform ball distribution to the conditional distribution of $y$. The fit uses the amortized input-convex network of the neural transport conformal paper. Only the conditioning input changes: it now includes the time gap.
The rank of a new change vector is
$$ r(y; x, g) = \big\| \nabla \psi^{*}(y; x, g) \big\|, \qquad s(y; x, g) = \frac{\nabla \psi^{*}(y; x, g)}{r(y; x, g)}, $$
where $\psi^{*}$ is the convex conjugate of $\psi$, so its gradient is the inverse map from data back to the ball. Here $r$ is the center-outward rank and $s$ is the direction on the unit sphere.
The learned map is only approximately optimal, so we do not trust the rank's nominal uniformity. Instead, we compute $r$ on a held-out calibration set of normal subjects. We use one interval per subject, which keeps calibration points exchangeable across subjects. The conformal p-value of a test interval is its position among these calibration ranks. An alarm at level $\alpha$ then has a false-alarm rate of at most $\alpha$ for any new normal subject exchangeable with the calibration set, whatever the quality of the transport fit. Fit quality affects only power.
We analyze three things. First, power as a function of time gap, since short gaps give noisy changes and the conditioning should widen the contours accordingly. Second, whether sign vectors at alarm cluster into interpretable progression directions, for example hippocampus-led versus ventricle-led. We check such clusters against later diagnosis rather than presenting them as subtypes by assertion. Third, robustness to scanner-harmonization outliers. That paper flags this weakness of global transport, so we compare against spatial ranks.
Baselines
The naive reference point is the per-subject least-squares slope of hippocampal volume, thresholded at the 5th percentile of normal subjects' slopes. This is what a clinic could compute today. It is the comparison point against which the conformal bands paper reports its 17.5% gain.
The strongest existing clinical method is the Rate of Change Bound from Uncertainty-Calibrated Prediction of Randomly-Timed Biomarker Trajectories with Conformal Bands. We run it in two forms. The first applies it to each biomarker separately and reports the best single biomarker. The second combines biomarkers with a Bonferroni correction so that it matches our joint false-alarm level. Its reported result is 17.5% more MCI progressors identified than slope methods, at nominal 90% coverage. Beating it means that at the same empirical false-alarm rate on held-out normal subjects, our method detects more eventual converters. The bootstrap confidence interval on the sensitivity difference must exclude zero.
The strongest multivariate competitor is a conformalized conditional Mahalanobis score. A network predicts the conditional mean and full covariance of $y$ from $(x, g)$, and the squared whitened residual is calibrated on the same normal subjects. It captures correlation and conditional spread at almost no cost, so it is where a skeptic expects our gains to disappear.
We also run a scalar max-norm score, the largest standardized residual across biomarkers. This is the most direct multivariate extension of the conformal bands construction. OT-CP+ from Optimal transport-based conformal prediction and GRCP from Geometric Conformal Prediction with Spatial Ranks and Multivariate Quantiles serve as off-the-shelf multivariate conformal methods. In both, the time gap enters only as an extra covariate.
Every method gets the same things. It sees the same subject-level train, calibration and test splits, and the same predictors including the time gap. It gets the same hyperparameter search budget, set in GPU-hours, with selection by a proper scoring rule on validation normal subjects and never on converters.
Three ablations isolate the key ingredients. The first removes the time gap from the conditioning. The second restricts the potential to a quadratic, which makes the transport linear and the rank equivalent to a Gaussian Mahalanobis score. This directly tests whether non-elliptical shape is what matters. The third swaps the ball reference for a hypercube. No multivariate conformal result on ADNI or BLSA was found in the research file for any of these methods, so all multivariate numbers will be new.
Why not just…?
Why not just run univariate conformal bands on each biomarker and Bonferroni-correct? This controls joint false alarms, but it ignores correlation. Hippocampal and entorhinal loss move together, so a patient with moderate joint loss may cross no single corrected threshold. On static benchmarks, Neural Optimal Transport Meets Multivariate Conformal Prediction shows coordinate-wise boxes are substantially larger than transport regions at equal coverage. Whether this becomes lost sensitivity in clinical data is exactly what the Bonferroni baseline measures.
Why not just conformalize a Mahalanobis distance with a learned conditional covariance? This is the real threat. If normal change is close to Gaussian given baseline and time gap, it will match our method at a fraction of the complexity. Our evidence that it can fail is indirect. Optimal transport-based conformal prediction reports poor worst-set coverage for ellipsoids, and the neural transport paper reports large volume losses on curved shapes. Neither result is clinical. The quadratic-potential ablation settles the question head-on, together with a pre-registered ellipticity test on the conditional residuals of normal subjects. If the full potential does not beat the quadratic one, we will report that longitudinal atrophy is effectively elliptical and recommend the ellipsoid.
Why not just collapse the change vector to one standardized norm? A max-norm or $\ell_2$ score keeps the guarantee but throws away direction. Geometric Conformal Prediction with Spatial Ranks and Multivariate Quantiles shows that scalarization gives oversized, unoriented regions when the spread is anisotropic, that is, much wider along some directions than others. The scalar score also cannot say which biomarkers drive an alarm. We measure the power cost directly. The interpretability cost shows up in the sign-versus-diagnosis analysis.
Why not just use discrete transport ranks (OT-CP+) as they are? Discrete transport scales cubically with calibration size. It handles the time gap only through nearest-neighbor localization, and short-gap intervals are sparse. We expect it to lose power on unusual time gaps. The stratified-by-gap analysis will show whether that is true.
Proposed Experiments
The primary experiment uses the harmonized ADNI and BLSA cohort described in the conformal bands paper. Splits are at the subject level: normal subjects go to training and calibration, and a held-out set of normal subjects plus MCI subjects forms the test set. Validity is empirical false-alarm rate on held-out normal subjects, at $\alpha = 0.05$ and $0.10$. It is reported overall and within the Mondrian strata of sex and APOE4 that the conformal bands paper already uses, together with worst-slab coverage. Power is sensitivity for MCI subjects who later convert, with conversion defined clinically and independently of the biomarkers. We also report the partial area under the ROC curve at low false-alarm rates, and median lead time between the first alarm and clinical conversion. As a sanity check, we report joint prediction-region volume per unit of time gap. The retinal-thickness data used by Unscrambling disease progression at scale is cross-sectional, so we do not use it for patient-level alarms.
The decisive checks are three, all pre-registered. First, every method must hold its false-alarm rate on held-out normal subjects, including within APOE4 strata. A method that gains power by breaking validity is disqualified. Second, the full transport rank must beat the quadratic-potential ablation and the conditional Mahalanobis baseline in converter sensitivity at matched false-alarm rate. If it does not, the core claim fails. Third, dropping the time gap from the conditioning must hurt most on short-gap intervals. If it does not, conditioning is not doing the work we attribute to it.
A secondary, exploratory check asks whether sign directions at the first alarm predict time to conversion better than the rank alone. This tests whether the "sign" half of the rank-and-sign pair carries clinical information.
References
- Monge–Kantorovich depth, quantiles, ranks and signs
- Vector quantile regression: An optimal transport approach
- Neural Optimal Transport Meets Multivariate Conformal Prediction
- Optimal transport-based conformal prediction
- Geometric Conformal Prediction with Spatial Ranks and Multivariate Quantiles
- Uncertainty-Calibrated Prediction of Randomly-Timed Biomarker Trajectories with Conformal Bands
- Unscrambling disease progression at scale: fast inference of event permutations with optimal transport
- DeltaNet: Conditional Medical Report Generation for COVID-19 Diagnosis
- Learning to Exploit Temporal Structure for Biomedical Vision-Language Processing
- CoCa-CXR: Contrastive Captioners Learn Strong Temporal Structures for Chest X-Ray Vision-Language Understanding