Relapses as Jumps: Pathway-Informed Continuous-Time Models of MS Progression

Two research communities studying multiple sclerosis (MS) have almost no contact. Immunologists have identified specific molecular switches inside T cells — Nur77, an internal brake on activation whose loss leaves cells chronically switched on; ANKRD55, a throttle that amplifies T-cell inflammation; let-7f-5p, a small regulatory RNA that normally suppresses the STAT3 gene and blocks inflammatory Th17 cells — that are dysregulated in patients compared to healthy controls, but almost always measured once, at a single clinic visit. Meanwhile, the machine-learning community has built increasingly good models of how clinical disability unfolds over time in MS registries, but these models see only macroscopic scores: walking speed, hand dexterity, a clinician-rated disability scale. Neither side can answer the question that matters clinically: does a patient's molecular immune profile, tracked over time, tell us anything about when the next relapse will strike?

This proposal argues that the obstacle is not just missing data but a wrong mathematical assumption. The strongest longitudinal MS models assume the patient's underlying health state evolves smoothly and continuously, yet MS is punctuated by relapses — sudden, discrete inflammatory attacks. We propose a continuous-time model whose latent state evolves smoothly between relapses but jumps discontinuously at relapse times, with the jump probability (a hazard rate) driven by a small, interpretable set of immune-pathway state variables built from the molecular checkpoints the immunology literature has already validated. The model is a direct extension of the Neural CDE architecture that currently holds the best published numbers on the MSOAC progression benchmark, and its two ingredients — jumps and pathway-structured latent dimensions — each fill a gap the existing papers explicitly name as open.

Core Claim

The claim has two parts, one architectural and one substantive. Architecturally: a continuous-time sequence model whose latent dynamics combine a smooth Neural CDE drift with an explicit marked point process for relapses will predict disability progression better than smooth-only continuous models, especially for patients whose histories contain sharp spikes — the exact subpopulation where the current best models struggle. Substantively: the latent state that drives relapse hazard can be structured so that individual dimensions correspond to named immune pathways (Nur77-mediated braking, STAT3/Th17 drive, innate nucleic-acid sensing), making the hazard interpretable — a clinician can ask "which pathway is pushing this patient toward a relapse?" — and making the model testable against mechanistic immunology rather than just against prediction metrics.

Formally, let $X_t$ be the observed biomarker path up to time $t$ (irregularly sampled clinical scores, and where available, molecular markers). The latent state $z_t \in \mathbb{R}^k$, with one dimension per pathway, evolves as a Neural CDE — a network whose hidden state changes in proportion to incoming observations, $dz_t = f_\theta(z_t)\,dX_t$ — and relapses arrive as a point process with intensity (instantaneous risk) $\lambda_t = g_\theta(z_t)$. At a relapse time $\tau$, the latent state jumps, $z_{\tau^+} = z_{\tau^-} + W_\theta m_\tau$, where $m_\tau$ is a learned relapse "mark" encoding severity and affected systems, and then relaxes back under the smooth dynamics. All symbols beyond $X_t$, $z_t$, $\lambda_t$, and $m_\tau$ are network parameters to be learned; the structure, not the specific parameterization, is the contribution.

What Prior Work Leaves Open

Three papers define the frontier this proposal pushes against. De Brouwer et al. benchmarked five architectures on the MSBase registry and showed that longitudinal models beat a static snapshot Random Forest substantially — ROC-AUC of 0.86 versus 0.79 for predicting two-year confirmed disability progression, with precision at 80% recall rising from 29% to 42% — but their input features were exclusively clinical, and they explicitly note that their best continuous model (GRU-ODE-Bayes) enforces a smooth latent trajectory and therefore "struggles to capture acute, sudden, discrete relapse spikes" (Longitudinal modeling of MS patient trajectories improves predictions of disability progression). They also observed a striking clinical signal that smooth models cannot represent: temporary drops in disability frequently precede permanent progression. A jump-relaxation model is the natural shape for that phenomenon.

Norcliffe et al. benchmarked continuous-time models on the MSOAC placebo dataset and found that Neural CDEs beat the best discrete baseline (a Temporal Convolutional Network) in 10 of 16 tasks, achieving an 18–24 month EDSS regression RMSE of 1.587 versus 1.666 for the TCN, and an AUPRC of 0.658 for classifying severe disability (EDSS > 5) at two years (Longitudinal Modeling of Multiple Sclerosis using Continuous Time Models; Benchmarking Continuous Time Models for Predicting Multiple Sclerosis Progression). Two of their findings constrain our design. First, they document that Neural CDE dynamics become numerically unstable with exploding gradients at higher feature dimensions, which forced them down to four performance measures plus demographics — this is why we route molecular panels through a small pathway-structured latent space ($k \approx 4$–6) rather than feeding them in raw. Second, their ablation showed that feature standardization, not interpolation sophistication, dominates stability — without it, RMSE nearly tripled — so we adopt their preprocessing exactly and spend no effort on interpolation refinements.

Schulam and Arora's Disease Trajectory Maps showed on scleroderma cohorts that non-linear latent trajectory models can discover hidden subgroups ("late decliners" whose lung function stays flat for years before collapsing) and associate with outcomes more strongly than Linear Mixed Models (p = 0.002 versus p = 0.013 for pulmonary arterial hypertension), but their model is strictly univariate and tracks end-organ scores, never immune markers (Disease Trajectory Maps). Our proposal builds the multivariate, mechanism-bearing version their limitations section asks for. The Sow and Oxtoby review of data-driven disease progression modelling confirms the pattern across the field: progression models map organ damage sequences but have never incorporated post-transcriptional networks or T-cell lineage plasticity, and they assume a single underlying pathology (Data-Driven Disease Progression Modelling).

The immunology side supplies the pathway structure. Cappelletti et al. showed Nur77 braking is impaired in MS patients' CD4+ T cells (Quantitative proteomics reveals protein dysregulation during T cell activation in multiple sclerosis patients compared to healthy controls); Brorson et al. showed circulating CD8+ T cells are transcriptionally primed before active relapse — evidence that a measurable pre-relapse state exists (CD8<sup>+</sup> T cell gene expression analysis identifies differentially expressed genes between multiple sclerosis patients and healthy controls); Wu et al. showed ANKRD55 drives T-cell neuroinflammation and that disrupting it ameliorates disease (ANKRD55 is a key regulator of T cell inflammation in multiple sclerosis); Wang et al. established the let-7f-5p/STAT3/Th17 axis (Let-7f-5p suppresses Th17 differentiation via targeting STAT3 in multiple sclerosis); and Fukui et al. established TLR7 sensing as the SLE-relevant innate trigger (Anti-human TLR7 antibody for therapeutic intervention in systemic lupus erythematosus). None of these were ever measured serially in humans; that is the substantive gap. Feasibility of serial immune monitoring in blood is supported by Sun et al., who showed absolute lymphocyte and CD8+ counts at fixed post-treatment timepoints carry survival signal in lung cancer (Effectiveness and mechanisms of lymphocytes at different time points in predicting consolidation immunotherapy...), and by Tian et al., whose 29-gene PBMC RT-qPCR panel in Takayasu's arteritis is the practical template for a compact, repeatable assay (Identification of the Association Between Toll-Like Receptors and T-Cell Activation in Takayasu’s Arteritis).

Research Program

The program has three stages. Stage one builds and validates the jump-augmented Neural CDE on existing clinical-only benchmarks, where relapse and disability data already exist: MSOAC (2,465 patients, performance measures plus EDSS at four horizons) and MSBase (6,807 densely sampled patients, plus the roughly 87% of the registry excluded by the six-visit filter, which we use as an out-of-distribution test of whether the jump model degrades more gracefully under sparsity than the smooth models). The innovation tested here is purely architectural: does adding a marked point-process relapse component to the CDE improve prediction, and does it specifically improve prediction for the spike-containing histories that smooth models mishandle? We also fit hazard calibration curves — does the model's predicted relapse risk actually track observed relapse timing?

Stage two builds the synthetic mechanistic benchmark. Because no longitudinal molecular-checkpoint dataset exists in the retrieved literature (no reported result found on Lacuna for serial measurement of Nur77, ANKRD55, or let-7f-5p in any human cohort), we simulate patient trajectories from generative dynamics whose parameters are anchored to the qualitative findings above: Nur77 braking strength decays, releasing latent activation; STAT3 drive accumulates when let-7f-5p is repressed; relapses fire stochastically when a weighted combination of pathway states crosses a threshold; disability accumulates as a mixture of slow drift and relapse jumps with partial recovery. The benchmark tests whether the model recovers the true latent pathway states and jump timing from noisy, irregularly sampled observations — the identifiability question that determines whether the interpretable hazard is a genuine measurement or a story. Simulation code and the generative parameters become a public benchmark, filling the same role for mechanism-aware trajectory models that the Carass et al. 82-scan longitudinal MS MRI resource fills for imaging (Longitudinal multiple sclerosis lesion segmentation data resource).

Stage three is a prospective molecular pilot: roughly 60 relapsing-remitting MS patients sampled at three-month intervals for two years with a compact PBMC panel (Nur77 and ANKRD55 protein, let-7f-5p abundance, STAT3 phosphorylation, TLR7 expression, and flow cytometry for Treg/Teff and CD8+ activation, following the assay design of Tian et al.). The pilot's endpoint is not prediction accuracy — the cohort is too small — but whether pathway-latent trajectories estimated from the molecular data correlate with the relapse hazard learned from the clinical benchmarks, i.e., whether the two halves of the model describe the same biology.

Baselines

The "do nothing" reference point is the static snapshot Random Forest using only the most recent visit's features, exactly as run by De Brouwer et al. on MSBase, where it achieved an ROC-AUC of 0.79 for two-year confirmed disability progression (Longitudinal modeling of MS patient trajectories...). It matters because it quantifies how much of the predictive signal is already in the last observation; any model must clear it to justify tracking anything over time.

The strongest existing method depends on the benchmark. On MSOAC it is the smooth Neural CDE of Norcliffe et al., which reported an 18–24 month EDSS regression RMSE of 1.587 and an AUPRC of 0.658 for EDSS > 5 at two years (Benchmarking Continuous Time Models...). On MSBase it is the GRU with time-aware inputs (GRU-TA) and the tensor-factorization hybrid (BPTF-SI-RF) of De Brouwer et al., which reported ROC-AUC 0.86 and precision of 42% at 80% recall (Longitudinal modeling of MS patient trajectories...). We will rerun both under our exact preprocessing (Norcliffe's standardization and pseudo-time batching, identical visit windows, identical horizon definitions) with the same tuning budget via nested cross-validation, and on the same hardware, so that the only difference is the jump component. Beating the smooth Neural CDE means a lower RMSE than 1.587 at 18–24 months on MSOAC and a higher AUPRC than 0.658 for EDSS > 5 at two years, with confidence intervals from patient-level bootstrap; beating GRU-TA on MSBase means ROC-AUC above 0.86 and precision above 42% at 80% recall, plus, critically, better calibration of predicted relapse timing.

The most important ablation isolates the pathway-structured latent space: the identical jump-augmented architecture but with an unconstrained learned embedding (or PCA of the input features) in place of the pathway-aligned $z_t$. A second ablation removes the jump component, leaving a standard smooth Neural CDE on the same features, which directly measures the value of the jump mechanism independent of any molecular input. A third ablation replaces the point-process hazard with a simple binary "relapse in next window" classifier head, testing whether the explicit temporal point-process structure earns its place or a cheaper discrete head suffices.

Why not just…?

Why not just add biomarker columns to an existing GRU-ODE model? This is the most reasonable shortcut and we treat it as a serious baseline rather than an argument to dismiss. Two pieces of evidence say it will fall short. De Brouwer et al. explicitly report that their continuous model's smooth latent trajectory cannot represent acute relapse spikes, so extra input features cannot fix a representational problem in the state dynamics. Norcliffe et al. document exploding gradients when Neural CDEs take higher-dimensional inputs, so naively appending a molecular panel would destabilize the exact architecture a reviewer would want to extend. Still, the honest position is that if the plain GRU-ODE-Bayes with biomarker features matches our jump model on MSBase-style prediction, the architectural claim dies; we will run this comparison directly, and the subgroup analysis on spike-containing histories is where we expect the difference to appear.

Why not just use a discrete sequence model (TCN or LSTM) with time bins? The file's own benchmark answers this: Norcliffe et al. found the Neural CDE beat the TCN in 10 of 16 tasks, with RMSE 1.587 versus 1.666 at the 18–24 month horizon, and showed that discrete models require binning and zero-imputation that distort irregular clinical timelines (Benchmarking Continuous Time Models...). The margin is real but modest, so the TCN stays in our baseline suite; if the jump-augmented model does not also beat the TCN, not just the smooth CDE, the proposal's advantage is too thin to pursue. One genuine threat remains: a discrete model can represent spikes trivially, so on the spiky-history subgroup it may outperform smooth continuous baselines without any point-process machinery — which is precisely why that subgroup analysis is decisive for us.

Why not just PCA the immune panel down to a few dimensions and feed a standard Neural CDE? This addresses Norcliffe's instability finding at a fraction of the complexity, and it is a real threat to the pathway-structured embedding specifically. We concede the point conditionally: the prediction-accuracy ablation (PCA embedding versus pathway embedding, same jump architecture) settles it empirically, and if PCA matches on accuracy, we drop the mechanistic embedding claim. What PCA cannot provide is the interpretability endpoint: the pilot's core question — whether the dimension driving relapse hazard corresponds to the Nur77/STAT3 biology the immunology papers established — requires dimensions that are committed to pathways by construction. If PCA wins on accuracy and the pathway model still wins on hazard–biology correlation in the pilot, the mechanistic structure survives as an analysis instrument rather than a predictive one, and we would report it that way.

Proposed Experiments and Decisive Checks

The first experiment benchmarks on MSOAC: all four prediction horizons (0–6, 6–12, 12–18, 18–24 months), reporting RMSE for EDSS regression and AUPRC for EDSS > 5, against the smooth Neural CDE (RMSE 1.587, AUPRC 0.658 at the longest horizons) and TCN (RMSE 1.666) as reported by Norcliffe et al. The pre-registered success criterion is beating both on the 18–24 month horizon with non-overlapping bootstrap intervals, plus a materially larger margin on the pre-specified subgroup of patients whose observed EDSS histories contain sharp fluctuations — the population where De Brouwer et al. found smooth dynamics fail and where temporary EDSS drops precede progression (Longitudinal modeling of MS patient trajectories...).

The second experiment uses MSBase in two tiers: the dense 6,807-patient cohort for comparison against the published ROC-AUC 0.86 and precision-at-80%-recall 42%, and the sparse tier — the patients excluded by the six-visit filter — as the generalization test, where the hypothesis is that the jump model's explicit relapse mechanism compensates for missing observations better than interpolation-based smooth models. No reported result exists on Lacuna for any model on the excluded sparse tier, so this comparison establishes a new reference point rather than beating a published one.

The third experiment is the synthetic mechanistic benchmark: recovery error for latent pathway states and jump timing, hazard calibration, and a negative control in which relapse timing is generated independent of the pathway states (the model must then fail to find a pathway–hazard link, which demonstrates the benchmark can distinguish signal from story). The fourth is the molecular pilot, whose decisive check is the correlation between the estimated pathway-latent trajectories and the relapse hazard learned from the clinical benchmarks, pre-registered with a null of no association.

The program closes with what would falsify the core claim. If the jump component adds nothing over the smooth Neural CDE on the spiky-history subgroup; if PCA embeddings match the pathway embedding on prediction; or if the synthetic benchmark shows the pathway states are not identifiable from realistically sparse observations — then the interpretable, mechanism-bearing trajectory model is not yet achievable, and the honest fallback is a purely architectural contribution (jump-augmented continuous models) plus a public benchmark that lets the field test the molecular question the moment serial checkpoint data exist. No reported result found on Lacuna for any longitudinal molecular-checkpoint model in autoimmune disease, which is exactly why the synthetic benchmark and the pilot are designed to make that gap measurable rather than merely lamented.