ELPAC: Endpoint-Anchored Latent Progression with Stage-Varying Multimodal Coordination

AI summary

Overview

Research area: Statistical machine learning for biomedicine — Bayesian latent-variable modeling of disease progression, applied to multimodal Alzheimer's disease and aging cohort data (stat.ML; biostatistics).

Technical level: Advanced. The paper assumes familiarity with Bayesian nonparametrics/hierarchies, factor analysis and identifiability theory, variational inference, and disease progression modeling.

Scope (one sentence): The paper introduces ELPAC, a Bayesian framework that jointly infers each subject's progression subtype, continuous stage, and stage-varying cross-modality covariance from single partially observed cross-sectional snapshots, validated on synthetic cohorts and a harmonized ADNI+NACC/SCAN cohort of 7,501 subjects.

What This Paper Is About

Aging cohorts collect brain imaging, plasma biomarkers, and cognitive scores, but most participants contribute only one partially observed snapshot — there is no repeated measurement or follow-up diagnosis to say where they sit along the disease course. The paper's goal is to infer three things at once from such snapshots: which progression subtype a person follows (including preserved healthy aging), how far along that subtype's trajectory they are, and how measurements across modalities co-vary differently at different stages.

The central observation is that these three quantities are mutually dependent — staging, subtype assignment, and multimodal coordination cannot be cleanly solved one at a time. A motivating example in the paper: two people both score 26 on the MMSE, but one is roughly 30% along an early-onset AD amyloid+tau trajectory (stage ≈ 0.3) while the other is far along healthy aging (stage ≈ 0.9).

Key Contributions

1. Endpoint-anchored staging as an inductive bias. Instead of growing trajectories forward from an uncertain early-disease origin, ELPAC integrates subtype trajectories backward from clinically reliable endpoint groups toward a shared healthy reference, along per-modality velocity components that encode monotone direction, biomarker inclusion, onset timing, and optional sequential-ordering priors.

2. A stage-varying coordination decomposition with conditional identifiability. Residual variation around the inferred stage axis is factorized into sparse "coordination axes" whose loadings reveal which multimodal programs operate in a subtype and whose activations reveal when they emerge. The paper proves (Theorem 1) that the loadings, activation functions, and idiosyncratic noise are uniquely determined given the covariance identified at two stages with pairwise-distinct activation ratios, and the inferred stage axis itself supplies the auxiliary variation needed for identifiability — a setting where prior identifiable factor-analysis work assumes the auxiliary variable is observed.

3. Scalable joint inference. A structured amortized variational scheme pairs a low-rank Gaussian over global parameters with subtype-conditional Gaussians on logit-stage per subject, using a collapsed evidence lower bound that marginalizes subtype exactly; per-subject likelihoods are computed via the Sylvester–Woodbury identity at cost O(RK² + K³), and new subjects are scored in a single forward pass.

4. Empirical validation on synthetic data and a large harmonized cohort. Synthetic benchmarks (100 replicate cohorts) test recoverability of the anchored stage and the dynamic factor decomposition; a harmonized ADNI+NACC/SCAN cohort (N = 7,501; NACC March 2026 data freeze) tests held-out generative fit, subtype and stage recovery, prediction-interval coverage, and biological interpretability.

Main Findings

• Synthetic latent recovery is strong and joint. At N = 1,000 over 100 replicates, ELPAC achieves NLL/cell 1.07 ± 0.00, BAcc 95.4 ± 0.2, Macro-F1 96.0 ± 0.1, stage-score Spearman ρ 96.2 ± 0.0, masked NLL 0.91 ± 0.00, and |Cog, r| 97.9 ± 0.1 (metrics scaled by 100). It leads on every metric tied to the latent stage axis; its AUC of 81.0 ± 2.3 is below XGBoost (75.9 ± 2.3 is lower — see note) and label propagation (77.3 ± 0.3), which directly use endpoint labels; those supervised baselines, however, post much lower stage-score ρ (95.6 ± 0.0 and 43.9 ± 0.7 respectively) and no generative fit.
• Endpoint anchoring carries the subtype signal. Removing anchors (the "no-anchor" ablation) drops BAcc by 46.5 percentage points and cognition alignment by 21.8 points, and worsens held-out likelihood by 2.8 nats, while stage Spearman ρ stays within 4 points of the full model — indicating that boundary supervision mainly orients the latent axis and identifies subtypes.
• Stage-varying covariance is essential to the density model. Replacing the factor covariance with a diagonal one costs +4.7 nats held-out likelihood, +0.38 nats masked reconstruction, and −26.7 points BAcc.
• A static factor model is not sufficient. A static-factor covariance variant nearly matches on subtype and stage (BAcc 95.5 ± 0.2, stage ρ 97.4 ± 0.0) and even on masked NLL (0.912 ± 0.002) and cognition alignment (98.7 ± 0.1), but is clearly worse on held-out NLL (19.38 ± 0.05 vs 19.27 ± 0.05) and on the stage-binned posterior predictive check (78.7 ± 0.5 vs 82.8 ± 0.4) — the stage-resolved variation is what the full model adds.
• Recovered structure is interpretable in simulation. The subtype posterior cleanly separates the unlabeled population, recovered mean trajectories and onset locations are close to ground truth, and stage-varying cross-modal correlations track the true pattern across stages.
• Real-cohort claims are qualitative in the provided text. The abstract and contributions state that ELPAC shows accurate latent-structure recovery, well-calibrated prediction intervals, and biologically interpretable multimodal coordination programs that differ by subtype on ADNI+NACC/SCAN; the numeric real-cohort tables are not included in the truncated content, so specific values for that cohort are not reported here.

Methodology in Plain English

Two interlocking pieces. The model has a "progression" piece and a "coordination" piece, fitted together so each improves the other.

Piece 1 — build trajectories backward from the endpoint. Each subtype has a trajectory of expected biomarker values as a function of stage s ∈ (0,1). Rather than starting at stage 0 and extrapolating forward, ELPAC starts at the clinically confirmed endpoint group (near s → 1), where labels are trustworthy, and integrates backward in "shells" toward a shared healthy origin. The velocity at each shell is factored into three interpretable parts: whether a biomarker is involved in that subtype (a Bernoulli inclusion indicator with a Beta prior), when it turns on (a sigmoid in stage with its own onset center and sharpness), and how strongly it moves (a positive magnitude from a B-spline basis, multiplied by a fixed per-modality sign such as +1 for accumulation or −1 for atrophy). A soft coupling potential pulls every subtype's integrated starting point toward one shared healthy reference. Subject-level subtype is a covariate-modulated categorical, and stage is logit-normal; anchors are pinned near the endpoint and healthy controls near the origin.

Piece 2 — decompose what is left over. After subtracting the mean trajectory, the remaining residual variation is modeled as a low-rank factor model with a diagonal noise term. Each factor is a "coordination axis": its loadings say which regions and modalities move together, and its activation — a smooth positive function of stage, expanded in a B-spline basis with a second-difference smoothness penalty — says when along progression that coupling is strong. Loadings are constrained lower-triangular with positive diagonal to fix rotation, and a multiplicative-gamma shrinkage cascade automatically controls how many factors are effectively used. Activations are centered (spline coefficients summing to zero) so a constant rescaling can be absorbed into the loadings.

Why the two pieces need each other. A standard factor model is only identified up to rotation unless something external breaks the tie. Here that something is the stage axis recovered by Piece 1: if you know the covariance at two different stages and the factors switch on at different rates, a generalized eigenproblem recovers the individual axes. In return, the resulting stage-specific covariance sharpens the subtype and stage posteriors through the likelihood.

Fitting. Inference is amortized variational: a low-rank-plus-diagonal Gaussian over all global parameters, and a small recognition network (shared across subjects) that outputs a Gaussian on logit-stage for each candidate subtype. The subtype is marginalized analytically inside the bound, so the objective is a collapsed ELBO optimized with Adam and single-sample reparameterized gradients. Missing biomarkers are handled by simply selecting the observed subvector and submatrix of the multivariate normal likelihood; if the auxiliary cognitive outcome is missing, its factor drops out of the likelihood. Prediction intervals come from Monte Carlo draws of global parameters combined with the stage posterior, which marginalizes to a C-component Gaussian mixture on the logit stage.

Why This Matters

Research impact. The paper reframes an either/or problem — estimate trajectories or estimate covariance — as a joint one, and supplies an identifiability argument showing that the latent disease axis can itself serve as the auxiliary variable that makes a dynamic factor model identifiable. That is a meaningful shift for identifiable representation learning, where the auxiliary covariate is usually assumed observed. It also offers a principled answer to asymmetric supervision: use the labels you trust (endpoints, healthy controls) to pin the axis, and let the noisy middle of the cohort be inferred rather than used as an anchor.

Real-world applications.

• Trial enrichment and stratification: identifying which pre-symptomatic or mildly impaired participants are far along a fast-moving subtype versus simply old, so trials can enroll people likely to progress within the study window.
• Patient-level staging from a single visit: producing a continuous stage estimate with a credible interval from one workup of imaging, plasma, and cognitive tests, rather than waiting for longitudinal follow-up.
• Biomarker panel design: the loadings and activations indicate which modality combinations carry information at which stage, feeding "Fisher modality informativeness" summaries that could guide which tests are worth ordering when.
• Subtype-aware monitoring: because coordination programs differ by subtype and emerge at different stages, monitoring schedules can be tailored — e.g., imaging earlier for a subtype whose imaging coupling turns on early.

Industry relevance. For pharma and diagnostics groups working on neurodegeneration, the practical value is a generative, uncertainty-aware model that consumes the messy, partially observed, mostly unlabeled data real cohorts actually contain, handles held-out modalities (masked NLL is reported explicitly), and deploys on new subjects in a single forward pass rather than requiring re-fitting.

Future Directions

• Quantitative real-cohort reporting. The provided content describes the ADNI+NACC/SCAN results qualitatively; the full numeric comparison of held-out fit, stage/MMSE correlation, AUC, and 90% prediction-interval coverage against SuStaIn, DEBM, Static-FA, MoPoE-VAE, HEALNet, XGBoost, and label propagation is the natural next thing to examine.
• Validation against longitudinal outcomes. The framework is built for cross-sectional snapshots; checking whether inferred stages predict future conversion or cognitive decline in subjects who do have follow-up would test whether the anchored axis is prognostically as well as descriptively valid.
• Sensitivity to the anchor definitions. Anchoring is the load-bearing assumption, and the ablation shows subtype recovery collapses without it. How the results change with different endpoint definitions, different anchor fractions, or mislabeled anchors is an open question; the paper's priors explicitly allow informative ordering priors to be swapped in when external evidence supports them, which invites a systematic study of that prior specification.
• Scaling and model criticism. The amortized scheme is argued to work at cohort scale, and reproducibility code for the synthetic experiments is released; extending the identifiability guarantees beyond the stated conditions (data dimension R ≥ 2K+1, non-constant activations, distinct activation ratios at two stages) to regimes where those fail would clarify when the recovered coordination axes should be trusted.

Target Audience

Best suited to advanced readers: researchers in biostatistics, statistical machine learning, and computational neuroscience working on disease progression modeling, latent-variable identifiability, or multimodal fusion. It will be most valuable to those specifically interested in Alzheimer's and aging cohorts, Bayesian hierarchical modeling with variational inference, or identifiable factor models with auxiliary structure. Clinicians and industry scientists will find the framing and the reported subtype-specific coordination programs accessible, but the method sections require comfort with the underlying statistical machinery. Beginners should start with the introduction's MMSE-26 example and Figure 1's workflow diagram.

Authors’ abstract

Aging cohorts increasingly combine imaging, plasma, and cognitive measurements, but most subjects provide only a partially observed cross-sectional snapshot, leaving latent stage, progression subtype, and stage-varying multimodal coordination to be inferred jointly. Existing methods estimate the cohort-level aging process, model covariance under observed covariates, or fuse modalities for prediction. However, they rarely target the joint problem of endpoint-anchored staging and progression-resolved residual coordination across modalities, even though subject staging, subtype assignment, and multimodal interplay are mutually dependent. We introduce ELPAC, a Bayesian framework with two interlocking constructions. The first integrates subtype trajectories backward from clinically reliable endpoint groups along biologically structured velocity components encoding modality-specific monotone direction and optional sequential ordering priors. The second factorizes residual multimodal variation around the inferred stage axis into sparse coordination axes whose loadings and activations expose which multimodal patterns are operative in each subtype and when they emerge, with stage-resolved variation supplying the auxiliary structure for conditional identifiability. We use a scalable amortized variational inference scheme that enables joint posterior inference at cohort scale with single-pass deployment on new subjects. On synthetic data and a harmonized ADNI+NACC/SCAN Alzheimer's cohort, ELPAC demonstrates accurate latent-structure recovery, well-calibrated prediction intervals, and biologically interpretable multimodal coordination programs that differ by subtype.