An event-based model reconstructs the order in which biomarkers turn abnormal during a chronic disease, such as which brain region shrinks first, second, and so on. It does this from single-visit snapshots of many patients, each caught at a different point in their illness. The method is valued clinically because its output is itself the explanation: a readable ordering of events, plus a chart of how confident the model is about each event's position. Its weakest setting is also its most important one. Rare diseases such as corticobasal syndrome and progressive supranuclear palsy yield cohorts of a few dozen patients. Those cohorts are usually skewed: control subjects and mid-stage patients dominate, and the earliest and latest stages are nearly empty. When no patient sits at a given stage, the data carry almost no information about the order of the events around that stage. The model then either reports an arbitrary ordering or reports so much uncertainty that nothing is left to interpret.
We propose to fill those empty regions with information borrowed from a large related cohort, and to let the target data decide how much to borrow. The target cohort's event ordering gets a prior centred on the ordering learned from the source cohort, and the prior's strength is estimated rather than fixed. We also replace the usual uniform assumption about how patients spread across stages with an explicit, estimated stage distribution. The claim we will test has two parts. First, in small skewed cohorts, ordering errors concentrate at poorly populated stages. Second, a learned-strength prior corrects those errors without imposing the source ordering where the target data disagree. The result should be orderings that are more accurate and uncertainty estimates that are honest.
Core Claim
The standard event-based model and its subtype-aware extension, SuStaIn, handle each cohort on its own. SuStaIn, from Uncovering the heterogeneity and temporal complexity of neurodegenerative diseases with Subtype and Stage Inference, jointly clusters patients into subtypes and orders events within each subtype. pySuStaIn implements it and uses sampling to produce the "positional variance diagram", the chart of each event's uncertain position. Both fit each cohort from scratch under a uniform prior over orderings.
That choice is harmless in ADNI-sized cohorts of several hundred subjects. It is costly in small ones. Saito et al. fitted SuStaIn to 25 corticobasal syndrome patients, 39 progressive supranuclear palsy patients and 50 controls, restricted the analysis to 10 brain regions, and observed wider positional uncertainty at the sparsely populated late stages. They also found much lower sensitivity for the smaller group: 0.68 for corticobasal syndrome against 1.00 for progressive supranuclear palsy. These results motivate the proposal. They show the regime exists and that the failure sits where our mechanism predicts.
Transfer from large to small cohorts has been demonstrated only for single time-to-event outcomes. Transfer Survival Boosting reweights source patients by their similarity to the target cohort. It raised the C-index on a small heart-failure study from 0.500 to 0.677, but it learns no ordering of intermediate events.
The research file records no method that transfers event orderings. It also records none that models skewed stage occupancy inside an event-based model. Our mechanism is a Mallows-type prior, meaning a probability distribution over orderings that decays with distance from a centre ordering:
$$ p(S \mid \theta) \propto \exp\big(-\theta \, d_K(S, S_{\mathrm{src}})\big). $$
Here $S$ is the target event ordering, $S_{\mathrm{src}}$ is the posterior mode ordering from the source cohort, $d_K$ is the Kendall tau distance (the number of event pairs the two orderings place in opposite order), and $\theta \ge 0$ is a concentration parameter with a weak hyperprior, sampled jointly with $S$. The value $\theta = 0$ recovers the standard model.
The per-patient likelihood marginalises over the latent stage $k$ with weights $\pi_k$ drawn from a Dirichlet prior, instead of the usual fixed $1/(N+1)$ for $N$ events.
The design principle is that borrowing supplies information, whereas reweighting only redistributes the information already present. Where target stages are populated, the likelihood dominates and $\theta$ has little effect. Where stages are empty, the prior fills the gap. Where the source is wrong, the posterior over $\theta$ should shrink toward zero.
Research Program
We will build the model on the mixture-model event-based code in pySuStaIn. Sampling over orderings stays feasible because rare-disease analyses use 10 to 20 regions. The continuous relaxation in vEBM, which optimises orderings through optimal transport, is a possible later route to voxel-level data. We will not depend on it here, because vEBM was evaluated on single trajectories in cohorts of 816 and 384 subjects, not in the small-sample regime.
The analysis has three parts. The first is a position-resolved error analysis. It measures ordering error as a function of how many target patients are estimated to occupy the stages adjacent to each event. This tests the mechanism directly, not just the aggregate score.
The second part extends the prior to SuStaIn. Each subtype gets its own centre ordering, taken from the matching source subtype, and its own concentration parameter. We will test whether this rescues minority subtypes. Scotton et al. report a progressive supranuclear palsy cohort of 426 patients in which only 17 have the subcortical variant, which shows how small such subtypes can be.
The third part treats interpretability as auditability. We will check whether the positional variance diagram is calibrated, meaning its 90% credible intervals contain the reference position about 90% of the time. A readable chart that is confidently wrong is worse than no chart. Where clinician-ranked event orders are available, we will also report Kendall tau agreement with experts, following Haddada et al.. They compared model feature rankings to clinician rankings in a static classifier; we apply the same measure to a temporal ordering.
Baselines
The naive reference point is the standard mixture-model event-based model, or SuStaIn with default settings, fitted to the small target cohort alone with a uniform prior over orderings and stages. This is what a clinical group would run today, and it is what Saito et al. ran. Every method receives the same regions, the same sampling budget (a fixed number of iterations and chains), and the same normal-versus-abnormal mixture fits.
The strongest competitor is pooled fitting. The source cohort and the small target cohort are concatenated, one model is fitted, and target patients are staged under the pooled ordering. It uses exactly the same extra data as our method at the same compute cost. It will likely win whenever source and target truly share an ordering. To beat it, our method must match it in shared-ordering conditions and clearly outperform it when the target ordering differs from the source. We will set a margin before running: lower Kendall distance to the reference ordering at the target's poorly populated positions, averaged over subsamples.
The second competitor is group-reweighted likelihood. Each target patient's contribution is weighted by the inverse frequency of their diagnostic group, the standard remedy for class imbalance. Its weights are tuned by cross-validation at the same budget.
Two ablations isolate our ingredients. The fixed-strength prior sets a large $\theta$, which amounts to copying the source ordering; it tests whether learning the prior's strength matters. The prior without estimated stage weights keeps the Mallows prior but uses uniform $\pi_k$; it tests whether modelling stage skew matters separately from borrowing.
On the clinical task, the strongest reported number in the file is Saito et al.'s 87.5% accuracy for separating corticobasal syndrome from progressive supranuclear palsy, with corticobasal syndrome sensitivity of 0.68 and 76.5% accuracy on an external dataset. For ordering-recovery error in small subsampled cohorts, no reported result was found on Lacuna for any event-based model, so we will produce that baseline ourselves.
Why not just…?
Why not just reweight the minority groups in the likelihood? Reweighting changes how much each patient counts. It cannot create evidence about the order of events at stages where no patient sits, so we expect it to add variance there, not remove error. It also inflates the apparent sample size of the rare groups, which should narrow the credible intervals without making them more accurate. The calibration check will show whether this happens. If reweighting matches our method on position-resolved error, the borrowing claim is refuted and we will report that result.
Why not just pool the source and target cohorts? This is the real threat, and the experiments are built around it. Pooling is likely optimal when the orderings truly match. The large cohort also swamps the target, so a genuinely different target ordering gets overwritten. The two conditions below were designed to separate these cases. In the shared-ordering condition, we expect a tie. In the mismatched condition, the Alzheimer's source ordering is hippocampus-led, whereas Saito et al. report a bottom-up midbrain-to-cortex pattern in progressive supranuclear palsy, so the ADNI prior points the wrong way. There we expect pooling to fail and the learned $\theta$ to fall toward zero. If $\theta$ does not fall, our method has no advantage over pooling, and the negative-transfer concern raised by Transfer Survival Boosting applies to us as well.
Why not just synthesise extra patients for the empty stages? Generating synthetic patients needs a model of the ordering to place them at the right stages, which is the very thing we are trying to estimate, so the argument is circular. There is also direct evidence against the approach in this regime. Limitations of Synthetic Data Generation in Specialized Data-Scarce Domains found that no generative model consistently beat a simple augmentation baseline in scarce specialised domains. It also found that classifiers trained on real data scored 11% to 45% higher recall on synthetic images than on real ones, meaning the synthetic samples were cleaner and easier than real cases. We will include only a duplication-based oversampling control, because synthetic patients would distort the stage estimates in the same way.
Proposed Experiments
Both conditions use controlled subsampling, so the reference ordering is known.
The shared-ordering condition uses ADNI as the source. OASIS-3 is the target; it is an independent aging cohort, and Chen et al. study whether Alzheimer's subtypes transfer to independent populations of this kind. The reference ordering is the full-cohort OASIS-3 fit. We will draw repeated target subsamples of 30, 50 and 100 subjects with controlled skew: in some the controls dominate, in others the late stages are removed.
The mismatched condition uses the same ADNI source ordering as a prior for progressive supranuclear palsy subsamples. These are sized to match Saito et al. and are drawn from the Scotton et al. cohort, whose full-cohort fit serves as the reference.
The primary metrics are Kendall distance to the reference ordering, reported separately for well-populated and poorly populated positions, and the coverage of the positional credible intervals. The posterior over $\theta$ is a diagnostic in its own right. It should be large in the first condition and near zero in the second.
For external validity, we will use held-out longitudinal ADNI follow-up. Scotton et al. report that 90% of patients stayed at the same stage or moved to a later one at follow-up, and 95% kept the same subtype; we will use those figures as a reference for plausible stability. We will check whether target-fitted staging from small subsamples reaches comparable rates, and whether it separates corticobasal syndrome from progressive supranuclear palsy at matched cohort sizes against Saito et al.'s 87.5%.
The decisive result is the pattern across the two conditions. The proposal succeeds if the learned-strength prior reduces error at poorly populated positions relative to the target-only model, ties pooling when the orderings match, and beats pooling when they do not, with calibrated intervals throughout.
References
- Uncovering the heterogeneity and temporal complexity of neurodegenerative diseases with Subtype and Stage Inference
- pySuStaIn: a Python implementation of the Subtype and Stage Inference algorithm
- Temporal Progression Patterns of Brain Atrophy in Corticobasal Syndrome and Progressive Supranuclear Palsy Revealed by Subtype and Stage Inference (SuStaIn)
- Boosting Transfer Learning with Survival Data from Heterogeneous Domains
- Unscrambling disease progression at scale: fast inference of event permutations with optimal transport
- Uncovering spatiotemporal patterns of atrophy in progressive supranuclear palsy using unsupervised machine learning
- Assessing the Interpretability of Machine Learning Models in Early Detection of Alzheimer's Disease
- Limitations of Synthetic Data Generation in Specialized Data-Scarce Domains
- Transferability of Alzheimer's disease progression subtypes to an independent population cohort