When patients with a degenerative brain disease are followed for years across hospitals, their scans are acquired on different machines. Each machine adds its own offset to measurements such as cortical thickness or regional brain volume. Today these offsets are removed by statistical harmonization tools that were built for one-scan-per-person data, and only afterwards does anyone fit a progression or risk model. This note proposes to reverse that order. Brain tissue loss does not reverse. So when a patient's measured cortical thickness jumps up at the exact visit where they moved to a new scanner, that jump is very likely scanner offset and not biology. A model that fits each patient's trajectory as a monotone (never-improving) curve and jointly estimates per-scanner offsets can use these impossible jumps to identify the offsets. It can do this even at small sites, where too few subjects exist to estimate offsets the usual way.
The opportunity matters because the two failure modes in this area pull against each other. Aggressive harmonization strips out real atrophy. Weak harmonization lets progression and risk models learn scanner signatures, and the cost is largest at small hospitals that want to use a model trained elsewhere. Monotonicity is usually treated as a constraint on the output of a risk model. Here it serves as a source of information about the measurement process. The downstream payoff is dynamic risk: a per-visit conversion-risk estimate that is not thrown off whenever the patient changes scanners.
Core Claim
We claim that a joint hierarchical model gives three things in the regime where some subjects change scanner during follow-up and some sites contribute fewer than about twenty subjects. The model has three parts: a latent monotone Gaussian process per brain feature on a subject-aligned disease timeline, per-site additive and multiplicative offsets with a prior shared across sites, and an unconstrained observation layer for clinical scores. Compared with the standard two-stage pipeline of cross-sectional harmonization followed by a monotone progression model, we expect it to recover scanner offsets more accurately, to preserve true atrophy rates better, and to produce per-visit progression risk that transfers better to a small held-out site. If the two-stage pipeline matches it in that regime, the claim is false.
A monotonic Gaussian process is a flexible random curve whose derivative is pushed to stay positive (or negative) through extra "virtual derivative observations" with a probit likelihood. Filippone and colleagues use this construction (Maurizio Filippone et al. (2019)), and MGPA builds on it (Monotonic Gaussian Process for spatio-temporal disease progression modeling in brain imaging data). Concretely, for feature $k$ of subject $i$ at visit $j$ acquired on scanner $s$, we model
$$ y_{ijk} = f_k\big(\tau_i(t_{ij})\big) + \gamma_{sk} + \delta_{sk}\,\varepsilon_{ijk}, $$
where $f_k$ is a monotone Gaussian process over disease time, $\tau_i(t) = \alpha_i t + \beta_i$ is a subject-specific linear time warp aligning onset and pace, $\gamma_{sk}$ and $\delta_{sk}$ are the scanner's additive offset and noise scale, and $\varepsilon_{ijk}$ is standard Gaussian noise. As in ComBat, $\gamma_{sk}$ and $\delta_{sk}$ receive hierarchical priors shared across sites, so a small site borrows strength from large ones. The difference from ComBat is that the biological term is a per-subject monotone curve rather than a population regression on age and diagnosis.
There are two sources of identification. The first is the ordinary between-site mean difference. The second is new: within-subject discontinuities at scanner transitions that a monotone $f_k$ cannot absorb. We deliberately use a linear warp, not MGPA's richer one, because MGPA reports identifiability and initialization problems. A small site cannot afford those problems.
Relation to Prior Work
Three groups of papers motivate the proposal.
The harmonization literature supplies the offset model and documents the gap. ComBat-GAM, compared in A Comparison of Methods to Harmonize Cortical Thickness Measurements Across Scanners and Sites, showed that modeling non-linear age effects during harmonization recovers far more true effects: 25 significant regions versus 6 for ComBat and 4 for mixed-effects models. That paper also notes the method is cross-sectional and ignores within-subject correlation across changing scanners, which is exactly the setting we target. ComBat++ (Detect and Correct Bias in Multi-Site Neuroimaging Datasets) showed that scanner signatures survive standard ComBat. It added principal components as substitute confounders, reducing dataset-of-origin prediction from 69% to 45% accuracy. Both methods protect biology only through measured covariates. Ours protects it through a structural assumption about each subject's trajectory, which needs no complete covariate annotation. DC-ComBat (A Decentralized ComBat Algorithm and Applications to Functional Network Connectivity) shows that the empirical-Bayes offset estimation can be distributed exactly across sites. We rely on this to argue the site-level part of our model is federable, but we do not study federation here.
The monotone progression literature supplies the trajectory machinery. RMLP (Learning the irreversible progression trajectory of Alzheimer's disease) penalizes decreases in risk between visits and reduced "violation ratios" relative to unconstrained models on ADNI. MGPA (Monotonic Gaussian Process for spatio-temporal disease progression modeling in brain imaging data) fits monotone Gaussian processes with time warping to ADNI imaging. The ODE-RGRU model (Deep Geometric Learning with Monotonicity Constraints for Alzheimer's Disease Progression) adds monotonicity losses to continuous-time recurrent forecasting on TADPOLE. All three treat site variation as absent or pre-removed. In our model, monotonicity is the lever that estimates site variation.
Two papers constrain the design. ConDOR (Conditional Diffusion with Ordinal Regression: Longitudinal Data Generation for Neurodegenerative Disease Studies) combines monotone ordinal progression with domain conditioning across ADNI and the smaller OASIS. It is the closest prior work touching monotonicity, multiple domains, and small data together. However, it generates synthetic sequences rather than correcting measurements. It also enforces monotonicity directly on disease stage, which the authors themselves flag as too rigid for clinical scores. That warning, echoed in MGPA, is why our monotone constraint applies only to imaging-derived tissue measures. Cognitive scores such as MMSE and ADAS-cog enter through a separate non-monotone observation model. The score is a link function of the latent anatomical state plus a visit-level Gaussian process that can go up or down. Transfer Survival Boosting (Boosting Transfer Learning with Survival Data from Heterogeneous Domains) handles small target cohorts by reweighting source patients. Its target-only Cox baseline reached a concordance of only 0.500 on one study, a reminder that small sites cannot fit risk models alone. TSB is static, however. We transfer instead by sharing the population curves $f_k$ and the site-offset prior, then fitting only a few per-site and per-subject parameters at the target.
On the dynamic-risk side, Optimal intensive care outcome prediction over time using machine learning shows, in intensive care data, that risk recomputed from accumulating observations beats a baseline snapshot. That study used day-2 cumulative AUC 0.895 versus APACHE-II's 0.836 at admission, and it pooled 11 ICUs without harmonization. We borrow the evaluation idea of scoring risk at each visit as data accumulate. We do not claim results for intensive care.
Research Program
The work has four stages.
The first stage builds the joint model and its inference. We will use sparse variational inference for the monotone Gaussian processes, with derivative pseudo-observations on a fixed grid of disease time, and conjugate-style updates for the site offsets. To keep a small site from being captured by a poor warp, we will fit the population curves $f_k$ on large sites first. Small-site offsets and subject warps are then fit with the curves held fixed, followed by a short joint refinement.
The second stage is an identifiability analysis. We will characterize when the offset $\gamma_{sk}$ is identified from transitions alone. Roughly, the detectable offset at a transition must exceed the plausible monotone change over the inter-visit gap plus noise. That rule gives a concrete prediction: offset recovery should improve with the number of transition subjects at a site and degrade as follow-up intervals lengthen.
The third stage adds the clinical-score observation layer and a per-visit risk read-out. The read-out is the posterior probability that the subject's latent disease time crosses the dementia threshold within a fixed horizon, updated at every visit.
The fourth stage is the small-site transfer protocol described in the experiments.
Baselines
The naive reference point is no harmonization. Raw features are fed to the same monotone Gaussian process progression model and to the same risk read-out. This shows how much scanner variance hurts at all.
The strongest existing competitor for the harmonization question is the two-stage pipeline: ComBat-GAM (A Comparison of Methods to Harmonize Cortical Thickness Measurements Across Scanners and Sites) with age and diagnosis protected, followed by our exact monotone Gaussian process model with offsets fixed to zero. We will also run a ComBat++ variant with principal-component substitute confounders (Detect and Correct Bias in Multi-Site Neuroimaging Datasets). The two-stage pipeline uses the same downstream model, features, and splits, so any gain is attributable to joint estimation. It is the strongest competitor because it already protects non-linear biology. Its only missing ingredient is the within-subject transition signal.
The strongest existing competitor for forecasting is the monotonicity-constrained ODE-RGRU on TADPOLE (Deep Geometric Learning with Monotonicity Constraints for Alzheimer's Disease Progression). It reports clinical-status mAUC 0.881±0.022, against 0.878±0.022 for DeepRNN and 0.842±0.035 for PLSTM-Z. We will rerun it with its published configuration and an equal hyperparameter search budget, on both raw and ComBat-GAM-harmonized inputs. We will also run RMLP (Learning the irreversible progression trajectory of Alzheimer's disease) as the simplest monotone risk model. The file reports no violation-ratio figure for RMLP, so we will measure it ourselves. No reported result was found on Lacuna for MGPA's numeric reconstruction error, so MGPA is compared only on our own synthetic benchmark.
The key ablation removes the transition signal. We refit the joint model after deleting post-transition visits from every subject who changed scanner. If gains vanish under this ablation, the mechanism is confirmed. If they persist, the gain is coming from something else.
Beating the baselines means four things. Lower offset-recovery error and smaller bias in recovered atrophy slope than the two-stage pipeline on synthetic data at small sites (5 to 20 subjects). Lower site-of-origin classification accuracy at equal or better preservation of diagnosis effects on real data. A held-out small-site forecasting mAUC whose confidence interval does not fall below the ODE-RGRU's own small-site result.
Why not just…?
Why not just run ComBat-GAM and then fit a monotone model? This is the real threat. With large sites and no scanner transitions, we expect the two-stage pipeline to match us, because both then estimate offsets from between-site differences. The proposal stands or falls on the small-site, transition-rich regime. The synthetic sweep and the transition-deletion ablation are designed to show whether a gap opens exactly there. If no gap appears, we will report that the two-stage pipeline suffices.
Why not just enforce monotonicity in the risk model and skip harmonization? A monotone fit faced with an upward scanner jump does not remove the jump. It flattens the trajectory around it, which underestimates the atrophy rate of every subject who changed scanner, and a downward jump is read as accelerated decline. RMLP-style penalties (Learning the irreversible progression trajectory of Alzheimer's disease) cannot distinguish the two cases. The no-harmonization baseline with the monotone model measures this slope bias directly on synthetic data with known truth.
Why not just add site as a random effect in a longitudinal mixed model? Mixed-effects models were the weakest method in the PTSD comparison: 4 regions versus 25 for ComBat-GAM (A Comparison of Methods to Harmonize Cortical Thickness Measurements Across Scanners and Sites). That comparison was cross-sectional, however, and a longitudinal mixed model with subject random slopes does see within-subject jumps. Our model is essentially this model with a monotone non-linear curve and time warping in place of linear slopes. We will include the linear version as a baseline to isolate whether monotone non-linearity matters beyond simply modeling subjects longitudinally.
Why not just remove site information adversarially? Adversarial site removal (Harmonizing Multi Site Neuroimaging Datasets) needs enough data per site to train a site discriminator, which small sites lack. It also corrects representations rather than measurements, so the corrected values cannot be checked against known atrophy rates. We consider it out of scope for small sites rather than a competitive baseline.
Proposed Experiments
The first experiment is semi-synthetic, because only there is the true offset known. We take ADNI regional volume and thickness trajectories from the TADPOLE subset, smooth each subject into a monotone reference curve, and reassign visits to simulated scanners. We inject known additive and multiplicative offsets at transition visits. We sweep subjects per site (5, 10, 20, 50), the fraction of subjects with a transition, offset size relative to annual atrophy, and visit spacing. Metrics are offset-recovery error, bias in recovered per-subject atrophy rate, and the rate of predicted reversals.
The second experiment uses real ADNI/TADPOLE data. We first identify subjects whose scanner or field strength changed across visits from the acquisition metadata. We then report site-of-origin classification accuracy after harmonization, following the ComBat++ protocol. We also report the within-subject reversal rate at transition visits, which should drop toward the rate at non-transition visits, and preservation of diagnosis-group differences. As an external check of cross-sectional site leakage, we apply the fitted site prior to the 17-dataset pool used in Detect and Correct Bias in Multi-Site Neuroimaging Datasets, where a random forest identified the source dataset with 71.5% accuracy before correction.
The third experiment, leave-one-site-out transfer, is decisive for the small-sample direction. We hold out ADNI sites, subsample each to 10 or 20 subjects, fit only that site's offsets and its subjects' warps, and score per-visit three-class status mAUC and conversion-horizon calibration against the baselines above. Cognitive scores enter through the non-monotone layer. A secondary check confirms that predicted MMSE can rise when observed MMSE rises, while the latent anatomy stays monotone.
Two decisive outcomes would refute the idea. First, the joint model fails to beat ComBat-GAM-then-monotone on offset recovery at 5 to 20 subjects per site when transitions are present. Second, the transition-deletion ablation erases the gain. Either result means monotonicity carries no harmonization information beyond what ComBat already extracts, and we would report that.
References
- Probabilistic disease progression modeling to characterize diagnostic uncertainty: Application to staging and prediction in Alzheimer's disease
- Monotonic Gaussian Process for spatio-temporal disease progression modeling in brain imaging data
- A comparison of methods to harmonize cortical thickness measurements across scanners and sites
- Detect and correct bias in multi-site neuroimaging datasets
- A Decentralized ComBat Algorithm and Applications to Functional Network Connectivity
- Learning the Irreversible Progression Trajectory of Alzheimer's Disease
- Deep Geometric Learning With Monotonicity Constraints for Alzheimer's Disease Progression
- Conditional Diffusion with Ordinal Regression: Longitudinal Data Generation for Neurodegenerative Disease Studies
- Boosting Transfer Learning with Survival Data from Heterogeneous Domains
- Optimal intensive care outcome prediction over time using machine learning
- Initiating PeriCBD to probe perinatal influences on neurodevelopment during 3–10 years in China