acceptodds
Under review as a conference paper at ICLR 2027

TRACE: Trajectory Reconstruction and Clustering from Sparse Evidence for Disease Subtyping and Staging

Abstract

Chronic progressive diseases often follow heterogeneous courses, with individuals differing in the pattern and timing of biomarker abnormalities. However, observed data are typically sparse and predominantly cross-sectional, with limited follow-up covering only short segments of disease progression. Reconstructing heterogeneous trajectories from such data requires disentangling variation along disease trajectories from subtype heterogeneity while inferring individuals' unknown subtype assignments and disease stages. Existing approaches may confound disease stage with subtype, scale poorly as the number of biomarkers increases, or fail to effectively integrate cross-sectional and longitudinal observations. We propose TRACE (Trajectory Reconstruction And Clustering from Sparse Evidence), a probabilistic framework that jointly reconstructs biomarker trajectories and infers disease subtypes and continuous stages. TRACE combines monotonic neural networks for flexible modeling and data-driven clustering of subtype-specific progression trajectories with temporal regularization derived from longitudinal observations when available. All model components are jointly optimized end to end using stochastic gradient descent, enabling efficient and scalable estimation. We evaluate TRACE on semi-simulated MRI biomarker data with known subtype and stage structures and demonstrate its ability to identify clinically meaningful subtypes and disease stages in real-world neurodegenerative disease datasets.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Disease Progression Modeling, Sparse Longitudinal Data, Biomarker Trajectories, Neuroimaging

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_fxRaOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The goal of the paper is to assign patients to disease subtypes and disease stages, as well as reconstruct biomarkers' progression trajectories, based on both cross-sectional and longitudinal datasets. The authors compared their method, TRACE, with other baselines using real-data inspired synthetic datasets and claimed performance improvement. The authors also applied TRACE to real-world datasets.

Critical strengths

Assigning patients to different subtypes and disease stages using longitudinal data is a very important task with considerable impact on medical applications. The current paper is among a very few to solve this problem using a statistical approach. The current paper is also one of the few (maybe the first) that reconstruct biomarker trajectories from cross-sectional datasets.

Critical weaknesses

I have two main concerns over the paper: 1) incomplete and unfair benchmarking, and 2) lack of empirical support.

Incomplete and unfair benchmarking

Incompleteness

  1. "Bayesian Event-Based Model for Disease Subtype and Stage Inference" (Hao & Austerweil, 2026) did exactly the same tasks as, and improved upon, SuStaIn, but was not included as one of the baselines. Considering that it performed better than SuStaIn, I don't think TRACE's results are convincing without it.

  2. The papers lacked strong baselines for the special case of K=1K = 1. For cross-sectional data, candidate baselines are

    • DEBM (Venkatraghavan et al., 2019, Disease progression timeline estimation for Alzheimer’s disease using discriminative event based modeling)
    • EBM, KDE (Firth et al., 2020, Sequences of cognitive decline in typical Alzheimer’s disease and posterior cortical atrophy estimated using a novel event-based model of disease progression)
    • SA-EBM (Hao et al., 2025, Stage-Aware Event-Based Modeling (SA-EBM) for Disease Progression)
    • TEMPO (Hao & Austerweil, 2026, TEMPO: Transformers for Temporal Disease Progression from Cross-Sectional Data)

    Considering that the first three methods can only process ordinal disease stages, TEMPO, which outputs continuous disease stages as, is an important baseline to compare with in terms of Stage ρ\rho when K=1K=1. For longitudinal datasets, a strong baseline is TEBM (Wijeratne et al., 2023). For all the above, you can focus on Stage ρ\rho along with K=1K=1.

  3. The trajectory reconstruction as shown in Figure 3 lacks baselines. You are just showing TRACE's results, without any comparison with results from competitive methods. As far as I know, SubLign and DPMoSt both can reconstruct biomarker trajectories using longitudinal data. I think it's important to use the metric of integrated squared error (ISE) to compare your result with these two methods.

  4. LN 415-420 mentioned how TRACE identified the best possible number of subtypes for the real-world data. However, in synthetic experiments, this ability of TRACE was never tested against other baseline models. I think this is an important miss. As far as I know, both BEBMS ("Bayesian Event-Based Model for Disease Subtype and Stage Inference") and SuStaIn can do this. I recommend the authors to generate datasets with different number of subtypes and benchmark your method against BEBMS and SuStaIn using Mean Square Error (MSE) regarding estimating the number of subtypes.

  5. What's the runtime and resource consumption comparison? You only had it in B.4 (Ln 901-917) but not anywhere else.

  6. You used s-SuStanIn, not SuStanIn itself. Could you prove that s-SuStanIn performs at least as well as SuStaIn on your tasks? I know the s-SuStanIn paper has the benchmark but I think it's important to demonstrate it using your data as well. Also, SuStanIn has two variants (GMM and KDE) depending on the assumption of the distribution of biomarker measurements. I don't think you tested both.

Unfairness

  1. Ln 690. Using uniform distribution to generate disease stage is unrealistic. For example, Donohue et al. (2014) ("Estimating long‐term multivariate progression from short‐term data") suggested that later-stage participants are underrepresented. The result might be a non-uniform empirical distribution for patient disease stages. Also, using uniform distribution may give TRACE unfair advantage in the experiment. I know SuStaIn and SubLign have the same assumption, but DPMoSt used a Gaussian prior and BEBMS had a symmetric Dirichlet prior. For fair comparisons, I recommend the authors use different priors for your synthetic datasets, for example, Gaussian, Beta, and uniform.

    In sum, 1) It's a modeling leakage as you are using the same distribution in your model (LN 144) as what you used for data generation (Ln 690); 2) It's impractical as empirical distribution of disease stages might not be uniform.

  2. Comparing TRACE with s-SuStaIn for longitudinal data is unfair as SuStaIn is designed for cross-sectional data. I understand the motivation, but I think it's better to make it explicit in the caption of table 1 (even though the point has been made in LN 81-83)

  3. It's unfair to compare the stages predicted by SuStaIn with yours. SuStaIn can only output ordinal stages. SuStaIn interprets "stage" as the number of biomarkers that have become pathological whereas your definition of "stage" is a 0-1 float indicating the severity of the disease. They are very different concepts. And since your ground truth is continuous (Ln 690) but SuStaIn only outputs ordinal, the comparison is not apple-to-apple at all. To make the comparison fair, at least for some synthetic datasets, you have to use ordinal stages as the ground truth.

  4. The experimental setting, or the data generation process, as described in Ln 677-680, is different from the underlying assumption of SuStaIn. SuStaIn assumes that all subtypes have the same ROIs, just with different orderings over ROIs. According to your data generation, I don't think the comparison is fair for SuStaIn. To make fair comparisons, I recommend the authors to ALSO test on datasets where all subtypes have the same set of ROIs.

    Also, what's your rationale of assuming that different subtypes of Alzheimer's affect different regions? And how did you get the 5 shared and 10 subtype-specific regions setting? How do you prove that it's a realistic configuration?

Lack of empirical evidence

The real-world data experiments' result lack empirical support. The authors found that K=4K=4 is the most possible number of subtypes. Any medical evidence supporting this finding?

Also, I don't think "These patterns broadly correspond to the clinical profiles (Tab. 2)" is convincing. First, Table 2 is based on your own assignment of subtypes, so it's a self-validating evaluation. Second, you are just describing the characteristics of the four subtypes you identified, without any previous medical literature supporting your findings. For example, which medical papers identified the four subtypes for Alzheimer as your finding? And how are your identified four subtypes corresponding to those subtypes?

Questions for authors
  1. LN 120: "where larger values indicate greater abnormality". In reality, this might not be case. For some biomarkers, larger values indicate smaller abnormality. Is TRACE able to accommodate this? If the number of biomarkers is huge, this preprocessing is complicated and prone to errors. Therefore, I recommend the authors to make TRACE able to handle the cases where smaller values might indicate greater abnormality as well.

  2. Your model assumes that biomarker values progress monotonically. However, in reality, this might not be the case. Reversion is possible (for example, Malek-Ahmadi, M. (2016). Reversion from mild cognitive impairment to normal cognition: a meta-analysis. Alzheimer Disease & Associated Disorders, 30(4), 324-330.). While your model is not able to accommodate this, SuStaIn can. This is because SuStaIn and BEBMS don't have assumptions about this monotonically. They just assume that biomarker measurements follow normal distributions. I don't think you covered this as one limitation of TRACE.

  3. Why did you use logistic trajectory for biomarker progression trajectories? Why is that realistic way to model how biomarkers progress in the real world? Any prior work supporting this assumption?

  4. Your method described in Section 2.2 and 2.3 is almost identical to the inference method in the paper of BEBMS. The only differences stem from how to handle disease stages: TRACE assumes continuous stages whereas both BEBMS and SuStaIn assume ordinal ones. I don't think the authors are blindly copying BEBMS's idea. If the authors used AI in the math formulation, it might be that the model got the idea from the BEBMS paper. If the authors didn't use AI in the formulation, it nonetheless indicate a lack of rigorous literature review.

  5. The B.4 experiments clearly shows that the only advantage of TRACE over SuStaIn, at least with that experimental setting, is smaller runtime, without performance gains. Is that true?

  6. For Section 3.2 (Semi-simulated data experiments), can the authors try different number of ROIs (in terms of total unique ROIs among all subtypes)? Say, 5, 10, and 100.

  7. The synthetic experiments used only patients data, without data from controls. I think it's important to include control data for two reasons. First, it can test the models' capability to handle the edge cases. If the model cannot distinguish patients from controls, it might indicate there is something wrong. Second, practically, people might be incorrectly diagnosed and the participants with the "patient" label might be healthy and those "healthy" might be actual patients. Could the authors accommodate this suggestion in the semi-simulated experiments?

7 comments
AuthorsOpenReview ↗

Response to Reviewer fxRa (Part 1)

We sincerely thank the reviewer for the time and care devoted to this detailed review. The comments have helped us improve many aspects of our work. We appreciate the opportunity to address these concerns and strengthen the manuscript.

We organize our responses using the following labels: I for incomplete benchmarking (the first numbered list in the review), U for unfair comparisons (the second numbered list), E for empirical support for the real-data results, and Q for the reviewer's questions. We are still working on revising the manuscripts and will upload the complete version before the discussion deadline.

I1: Include BEBMS as an additional baseline

We thank the reviewer for suggesting this comparison. We have added BEBMS to our comparative evaluations using its released implementation and recommended settings. The results are presented in the relevant responses below. We will also cite BEBMS in our revised manuscript.

I2: Missing staging-only (K=1) benchmark

We agree that setting K=1K=1 provides a useful test of staging accuracy. We therefore added benchmarks using two simulation generators: (1) our 96-ROI semi-simulated generator; and (2) SuStaIn's z-score generator, with F=10,25,50F=10,25,50 biomarkers. We compared TRACE with SuStaIn (z-score and GMM variants), s-SuStaIn, EBM, KDE-EBM, SA-EBM, DEBM, and BEBMS, using stage ρ\rho to assess agreement between inferred and ground-truth stages.

Results on SuStaIn's simulator.

MethodF = 10F = 25F = 50
TRACE0.971 ±\pm 0.0030.987 ±\pm 0.0020.994 ±\pm 0.001
SuStaIn (Z-score)0.971 ±\pm 0.0030.989 ±\pm 0.0020.993$^{\dagger}$
DEBM0.948 ±\pm 0.0070.970 ±\pm 0.0070.981 ±\pm 0.002
SA-EBM0.918 ±\pm 0.0150.958 ±\pm 0.0090.974 ±\pm 0.004
BEBMS0.917 ±\pm 0.0150.957 ±\pm 0.0100.976 ±\pm 0.003
SuStaIn (GMM)0.883 ±\pm 0.0180.920 ±\pm 0.0130.938 ±\pm 0.012
s-SuStaIn0.858 ±\pm 0.0220.902 ±\pm 0.0070.923 ±\pm 0.010
EBM0.814 ±\pm 0.0290.879 ±\pm 0.0080.890 ±\pm 0.012
KDE-EBM0.683 ±\pm 0.0730.740 ±\pm 0.0550.813 ±\pm 0.044

†^{\dagger} This entry is based on a single fit, which took 23.8 hours.

Results on our 96-ROI semi-simulated datasets with K=1K=1.

MethodStage ρ\rho
TRACE (λ\lambda = 30)0.888 ±\pm 0.050
DEBM0.863 ±\pm 0.035
TRACE (λ\lambda = 0)0.645 ±\pm 0.060
SA-EBM0.321 ±\pm 0.065
BEBMS0.307 ±\pm 0.061
EBM (GMM mixtures)0.297 ±\pm 0.045
SuStaIn (GMM)0.262 ±\pm 0.050
s-SuStaIn0.232 ±\pm 0.041
KDE-EBM0.111 ±\pm 0.053

On SuStaIn's z-score simulator, TRACE achieved staging accuracy close to that of z-score SuStaIn and higher mean accuracy than the other evaluated methods. On our 96-ROI semi-simulated datasets, TRACE with λ=30\lambda=30 achieved the highest mean stage ρ\rho (0.888), followed by DEBM (0.863). Removing the change reconstruction loss reduced TRACE's mean stage ρ\rho to 0.645, indicating that within-person changes also improve staging when there is only one subtype.

I3: Missing comparisons on the reconstructed trajectories

Figure 3 was designed to illustrate the effect of the change reconstruction loss. We appreciate the reviewer's suggestion and added comparisons with SubLign and DPMoSt on our 96-ROI semi-simulated datasets.

For each scan, we evaluated each method's fitted biomarker mean at its own inferred subtype and stage. We reversed any method-specific scaling to express all predictions in common z-score units and compared them with the noise-free ground-truth biomarker profile at that scan's true subtype and stage. We measured integrated squared error by averaging the squared differences across scans and ROIs.

As shown below, TRACE with λ=30\lambda=30 achieved the lowest mean reconstruction error for both affected and unaffected ROIs. Its low error on unaffected ROIs also indicates that it introduced little spurious abnormality in regions with no simulated disease effect.

MethodError, affected ROIsError, unaffected ROIs
TRACE (λ\lambda = 30)0.065 ±\pm 0.0190.005 ±\pm 0.003
TRACE (λ\lambda = 0)0.150 ±\pm 0.0260.046 ±\pm 0.010
SubLign0.544 ±\pm 0.0560.065 ±\pm 0.013
DPMoSt0.603 ±\pm 0.0570.100 ±\pm 0.014

(-> continue in Part 2)

AuthorsOpenReview ↗

Response to Reviewer fxRa (Part 2)

I4: Recovery of the number of subtypes was not benchmarked

We agree that recovery of the number of subtypes required a separate evaluation. Following the reviewer's suggestion, we used SuStaIn's simulator with F=25F=25 biomarkers and true K∈{2,3,4}K\in\{2,3,4\}, generating five datasets for each value of KK. We evaluated candidate values from K=1K=1 to K=5K=5 and measured the mean absolute error (MAE) and mean squared error (MSE) between the selected and true numbers of subtypes.

We first applied the same CVIC-based selection rule to SuStaIn, BEBMS, and TRACE, using the rule implemented in the BEBMS package: a larger KK is preferred only when it reduces CVIC by more than 6.

For TRACE, we also evaluated the joint selection rule used in our real-data analysis, which formalizes the combination of held-out likelihood, assignment stability, and subtype proportions. Using five-fold cross-validation with 10 initializations per fold, we started at K=1K=1 and accepted K+1K+1 only when all three conditions held: (1) total held-out log-likelihood improved in at least 8 of the 10 initializations; (2) the mean pairwise ARI between out-of-fold participant assignments across initializations was at least 0.60; and (3) every subtype contained more than 10% of all participants in every initialization.

Method and selection ruleMAEMSE
BEBMS, CVIC1.031.70
SuStaIn, CVIC1.332.73
TRACE, CVIC1.402.53
TRACE, joint (rule used in our real data experiments)0.501.07

Although the joint rule yielded lower MAE and MSE in this evaluation, both selection rules depend on prespecified thresholds. The CVIC rule uses a minimum improvement of 6, while our joint rule uses thresholds for likelihood improvement, assignment stability, and subtype size. Selecting these thresholds is challenging in real-world data because the true number of subtypes is unknown and no single metric can reliably identify the appropriate number of clusters across settings. This uncertainty is a broader challenge in clustering studies and and we admit it still remains as a limitation of our method.

I5: Runtime and resources comparison

We have expanded the original B.4 benchmark to report fitting time and memory usage on the main 96-ROI semi-simulated task, as well as scaling with the number of biomarkers on datasets generated using SuStaIn's simulator. We present these two evaluations below.

(a) Main 96-ROI task. We measured fitting time and memory on the main semi-simulated datasets, with each fit limited to eight CPU cores. Peak RAM is the peak resident memory of the whole process tree. The ARI and stage ρ\rho columns summarize subtyping and staging performance for comparison with computational cost. TRACE required approximately five minutes per fit, offering the best performance at a computational cost comparable to several of the evaluated methods.

MethodFit timePeak RAM (GB)ARIStage ρ\rho
TRACE311 ±\pm 13 s0.890.9030.878
s-SuStaIn305 ±\pm 30 s1.610.5690.317
DPMoSt73 ±\pm 1 s0.710.1720.338
Surreal-GAN435 ±\pm 49 s0.690.4810.817
SubLign105 ±\pm 2 s1.290.1470.267
BEBMS~3.7 h2.900.4590.322

(b) Scaling with the number of biomarkers FF. We also evaluated TRACE and other event-based models on cross-sectional datasets generated using SuStaIn's simulator with increasing numbers of biomarkers. The reported slope is obtained by least-squares regression of log median fitting time on log FF, using the available entries from F=10F=10 to F=400F=400 for each method. A slope of 1 corresponds to linear growth in fitting time. TRACE's empirical slope of 0.98 indicates approximately linear scaling over this tested range, whereas the measured costs of BEBMS and SuStaIn increased more steeply.

MethodF = 10F = 25F = 50F = 100F = 400Slope
TRACE19 s28 s43 s80 s6 min0.98
DPMoSt13 s22 s33 s58 s3 min0.78
Surreal-GAN91 s103 s2 min3 min5 min0.41
s-SuStaIn3 s36 s90 s4 min23 min1.33
BEBMS2 min8 min28 min100 min–1.78
SuStaIn (GMM)2 min14 min99 min9.8 h–2.70
SuStaIn (Z-score)32 min11.3 h––––
Dash indicates unavailable result due to its long fitting time.

(-> continue in Part 3)

AuthorsOpenReview ↗

Response to Reviewer fxRa (Part 3)

I6: s-SuStaIn was used instead of original SuStaIn (GMM and KDE)

We used s-SuStaIn in the main 96-ROI benchmark because it reduces computational cost by grouping biomarkers into a smaller number of clusters. In contrast, the cost of fitting the original SuStaIn models increases steeply as the number of biomarkers grows. The runtime results in I5 illustrate this limitation. Biomarker clustering in s-SuStaIn can sacrifice information and affect recovery accuracy, so we agree that s-SuStaIn alone does not fully represent the performance of the original SuStaIn variants. Improving scalability while retaining flexible trajectory modelling is one motivation of TRACE. To provide a broader comparison, we have added benchmarks using SuStaIn's simulator that include s-SuStaIn, as well as the GMM, KDE, and z-score variants of SuStaIn, as reported in U3.

U1: The uniform stage prior is shared by the generator and the model

In the original simulation, baseline stages were sampled uniformly, while stages at subsequent visits followed participant-specific progression and were not sampled from a uniform distribution. Nevertheless, we agree that the uniform baseline distribution may favor TRACE's uniform stage prior. To ensure fairer comparison, we regenerated the 96-ROI task with three alternative baseline-stage distributions while retaining the uniform stage prior in TRACE:

  • Beta(2,5): an early-skewed distribution with fewer observations at later stages.
  • Truncated Gaussian: N(0.5,0.22)N(0.5,0.2^2), truncated to [0,1][0,1] before rescaling to the baseline-stage range.
  • Symmetric Dirichlet mixture: 10 equal-width stage bins with probabilities drawn from Dirichlet⁡(1,…,1)\operatorname{Dirichlet}(1,\ldots,1) separately for each dataset, and uniform sampling within each bin. This produces a different, irregular stage distribution in each dataset.

The results are shown as follows, with each table entry reports showing ARI / stage ρ\rho across datasets. TRACE with λ=30\lambda=30 achieved higher mean ARI and stage ρ\rho than s-SuStaIn, DPMoSt, and BEBMS under every baseline stage distribution tested. These results suggest that its advantage is not restricted to uniformly distributed baseline stages.

Baseline stage distributionTRACE (λ\lambda=30)TRACE (λ\lambda=0)s-SuStaInDPMoStBEBMS
Uniform0.903 / 0.8780.832 / 0.6510.569 / 0.3170.172 / 0.3380.459 / 0.322
Beta(2,5), early-skewed0.782 / 0.8390.433 / 0.5050.283 / 0.1170.019 / 0.4190.118 / 0.132
Truncated Gaussian0.925 / 0.8140.875 / 0.5370.769 / 0.1440.200 / 0.3970.558 / 0.201
Symmetric Dirichlet0.892 / 0.8520.813 / 0.6140.613 / 0.2360.243 / 0.3640.439 / 0.263

U2: Comparing TRACE with s-SuStaIn for longitudinal data is unfair as SuStaIn is designed for cross-sectional data.

We thank the reviewer for raising this concern. We clarify that Table 1 evaluates the method under two settings: 1) a cross-sectional setting using only one scan per participant and 2) a longitudinal setting that additionally includes scans of follow-up visits. In the cross-sectional setting, TRACE is trained without the change reconstruction loss and achieves the highest mean ARI and stage accuracy ρ\rho among the evaluated methods. Thus, TRACE's performance advantage is also observed under a cross-sectional comparison and is not attributable solely to its use of longitudinal data.

Our motivation was to learn from datasets containing both cross-sectional observations and sparse longitudinal follow-up, as is common in many public datasets. TRACE accommodates cross-sectional scans alone and leverages within-person changes when repeated visits are available. We view this capacity to use sparse longitudinal information as an important contribution, offering an advantage over methods that treat visits as independent observations.

(-> continue in Part 4)

Reviewer_h2qFOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper jointly models disease subtypes and continuous stages using subtype-specific smooth monotonic trajectories, marginalization over subtype and stage, and a posterior-weighted change-reconstruction loss for repeated visits. Subject-level cross-validation guides subtype-number selection on ADNI/OASIS–3, followed by a four-pattern discovery-set fit and clinical associations. Additional experiments assess follow-up inference after baseline-only fitting, alternative trajectory families, and pySuStaIn-generated data.

Critical strengths
  1. The formulation fits the sparse mixed-observation setting. Joint optimization of stage integration, subtype mixtures, and monotonic networks avoids discrete event-order searches. The change loss provides substantial recovery improvements in the main simulation, and trajectory-family ablations support the SMM choice.
  2. Validation extends beyond discovery-set clustering. There is subject-level five-fold validation, follow-up inference after baseline-only fitting, and comparable recovery to SuStaIn on three pySuStaIn-generated datasets with reported reduced CPU time. These provide complementary evidence beyond clinical group significance alone.
Critical weaknesses
  1. The strongest results come from one semi-simulated dataset used for both parameter selection and evaluation. Independent simulations with different progression patterns are needed to establish robustness.
  2. The longitudinal loss uses changes between visits but does not explicitly model elapsed time or progression rate. The efficiency advantage on high-dimensional data also needs stronger direct comparisons.
Questions for authors
  1. Limited additional experiment: With λ_chg fixed do the main gains persist on independently generated 96-ROI simulation seeds? Please separate data-generation variability from initialization variability.
  2. Mechanism clarification: Do the authors agree that Equation 17 itself encodes neither direction nor interval length? Could an existing-data control using stage/baseline-matched cross-person pairs distinguish individual change information from generic reconstruction regularization?
  3. Efficiency clarification: Please report training time, memory, and stopping criteria for TRACE and s-SuStaIn/DPMoSt on the main 96-ROI task. Does the small-task speed advantage extend to the principal setting?
4 comments
AuthorsOpenReview ↗

Response to Reviewer h2qF (Part 1)

We sincerely thank the reviewer for their time and constructive comments. We address each concern below, using “W” and “Q” to refer to the numbered weaknesses and questions in the review, respectively.

1. W1 and Q1: Do the gains persist on independently generated semi-simulated datasets with λ\lambda fixed?

Yes. TRACE retained best performance across all reported metrics in both the cross-sectional and longitudinal settings. Following the reviewer's suggestion, we generated 10 independent 96-ROI semi-simulated datasets using different random seeds. We then refitted each method using three initializations per dataset. We averaged results across initializations within each dataset and report the mean and standard deviation across the 10 datasets. We have updated Table 1 in the manuscript with these results, as shown below.

MethodCross-sectional: ARI ↑\uparrowCross-sectional: Stage ρ\rho ↑\uparrowWith Longitudinal Visits: ARI ↑\uparrowWith Longitudinal Visits: Stage ρ\rho ↑\uparrowWith Longitudinal Visits: L-Stage ρ\rho ↑\uparrowWith Longitudinal Visits: JSSC ↑\uparrow
Progression-aware subtyping methods
SubLign––0.147±0.1010.147_{\pm 0.101}0.267±0.0850.267_{\pm 0.085}0.466±0.0300.466_{\pm 0.030}0.504±0.0820.504_{\pm 0.082}
Surreal-GAN0.064±0.0500.064_{\pm 0.050}0.295‾±0.092\underline{0.295}_{\pm 0.092}0.481±0.1140.481_{\pm 0.114}0.817‾±0.091\underline{0.817}_{\pm 0.091}0.559‾±0.132\underline{0.559}_{\pm 0.132}0.594‾±0.094\underline{0.594}_{\pm 0.094}
s-SuStaIn0.306‾±0.096\underline{0.306}_{\pm 0.096}0.281±0.1040.281_{\pm 0.104}0.569‾±0.083\underline{0.569}_{\pm 0.083}0.317±0.0960.317_{\pm 0.096}0.224±0.0300.224_{\pm 0.030}0.445±0.0340.445_{\pm 0.034}
DPMoSt0.057±0.0410.057_{\pm 0.041}0.148±0.0520.148_{\pm 0.052}0.172±0.0930.172_{\pm 0.093}0.338±0.0330.338_{\pm 0.033}0.466±0.0300.466_{\pm 0.030}0.510±0.0790.510_{\pm 0.079}
BEBMS0.118±0.1050.118_{\pm 0.105}0.292±0.0820.292_{\pm 0.082}0.459±0.1970.459_{\pm 0.197}0.322±0.1000.322_{\pm 0.100}0.253±0.0710.253_{\pm 0.071}0.413±0.0970.413_{\pm 0.097}
TRACE (Ours)0.365±0.146\mathbf{0.365}_{\pm 0.146}0.639±0.121\mathbf{0.639}_{\pm 0.121}0.903±0.023\mathbf{0.903}_{\pm 0.023}0.878±0.035\mathbf{0.878}_{\pm 0.035}0.799±0.044\mathbf{0.799}_{\pm 0.044}0.934±0.010\mathbf{0.934}_{\pm 0.010}
Cross-sectional subtyping methods
K-Means0.163±0.0750.163_{\pm 0.075}–0.392±0.2220.392_{\pm 0.222}–––
CHIMERA0.132±0.0670.132_{\pm 0.067}–0.218±0.0830.218_{\pm 0.083}–––
SmileGAN0.007±0.0140.007_{\pm 0.014}–0.000±0.0010.000_{\pm 0.001}–––

2. W2 (first part) and Q2: What information does the change reconstruction loss use, and do its benefits truly depend on within-person changes?

What information does the loss use? Equation 17 compares signed observed biomarker changes with signed differences between reconstructed biomarker profiles, so it captures both the magnitude and sign of change across ROIs. However, it does not use the elapsed time between visits. The loss provides a reconstruction constraint rather than a model of progression over calendar time (which is different from our defined latent disease stage).

The cross-sectional data likelihood objective evaluates each visit independently. The change reconstruction loss additionally encourages the two visits to be explained by a shared subtype (Eq. 15) and their observed biomarker changes to agree with differences between posterior-expected profiles along that subtype's trajectory (Eqs. 16–17). This constrains the shape of the trajectories using within-person changes. Taking differences also cancels stable additive participant-specific effects, which can help separate disease-related changes from persistent anatomical differences between participants.

Does the benefit depend on within-person changes? To test whether the benefit depends on valid within-person pairing, we conducted the following additional experiments: we changed only the follow-up observations used in the change reconstruction loss, while retaining every other model setting. Each follow-up was replaced by one from another participant, selected either randomly (Shuffled) or through matching (Matched). For matching, we selected donor pairs based on the earlier visits' inferred subtype and stage, a 10-component PCA representation of their 96-ROI biomarker profiles, and the log-transformed interval between visits. Results are reported as the mean and standard deviation across datasets.

Change-loss pairsARIStage ρ\rhoL-Stage ρ\rhoJSSC
Genuine within-person (TRACE)0.903 ±\pm 0.0230.878 ±\pm 0.0350.799 ±\pm 0.0440.934 ±\pm 0.010
None (λ\lambda = 0)0.832 ±\pm 0.0910.651 ±\pm 0.0550.652 ±\pm 0.0500.859 ±\pm 0.031
Shuffled cross-person follow-up0.441 ±\pm 0.1720.545 ±\pm 0.0880.451 ±\pm 0.0810.660 ±\pm 0.113
Matched cross-person follow-up0.048 ±\pm 0.0300.321 ±\pm 0.0270.404 ±\pm 0.0840.341 ±\pm 0.040

(-> continue in Part 2)

AuthorsOpenReview ↗

Response to Reviewer h2qF (Part 2)

These results support that the benefit of the change reconstruction loss truly depends on within-person pairing, rather than generic reconstruction regularization. Both cross-person pairing strategies performed substantially worse than omitting the loss entirely. Shuffled pairs also performed better than matched pairs. One possible explanation is that shuffled pairs retain larger differences in disease stage or subtype, whereas matching reduces these differences and may leave participant-specific anatomical variation as a larger component of the reconstruction target. Fitting this variation could encourage the model to group participants by individual anatomy rather than disease progression.

We will add these results to the revised manuscript.

3. W2 (second part) and Q3: Training time, memory, and stopping criteria on the main 96-ROI task

In response to this comment and related comments from other reviewers, we expanded original B.4 benchmark to assess both computational cost on the main 96-ROI task and scaling with the number of biomarkers. We report these two evaluations separately below.

(a) Main 96-ROI task. We measured fitting time and memory on the main semi-simulated datasets, with each fit limited to eight CPU cores. Peak RAM is the peak resident memory of the whole process tree. The ARI and stage ρ\rho columns summarize subtyping and staging performance for comparison with computational cost. TRACE used a fixed budget of 600 epochs without early stopping, and DPMoSt used 30 outer iterations with 30 inner iterations. s-SuStaIn used the updated configuration with 50 initializations, at most 200 iterations per initialization, a convergence tolerance of 10−810^{-8}, and 10 biomarker clusters with a minimum cluster size of five. TRACE required approximately five minutes per fit, offering the best performance at a computational cost comparable to several of the evaluated methods.

MethodFit timePeak RAM (GB)ARIStage ρ\rho
TRACE311 ±\pm 13 s0.890.9030.878
s-SuStaIn305 ±\pm 30 s1.610.5690.317
DPMoSt73 ±\pm 1 s0.710.1720.338
Surreal-GAN435 ±\pm 49 s0.690.4810.817
SubLign105 ±\pm 2 s1.290.1470.267
BEBMS~3.7 h2.900.4590.322

(b) Scaling with the number of biomarkers FF. We also evaluated TRACE and other event-based methods on cross-sectional datasets generated using SuStaIn's simulator with increasing numbers of biomarkers. The reported slope is obtained by least-squares regression of log median fitting time on log FF, using the available entries from F=10F=10 to F=400F=400 for each method. A slope of 1 corresponds to linear growth in fitting time. TRACE's empirical slope of 0.98 indicates approximately linear scaling over this tested range, whereas the measured costs of BEBMS and SuStaIn increased more steeply.

MethodF = 10F = 25F = 50F = 100F = 400Slope
TRACE19 s28 s43 s80 s6 min0.98
DPMoSt13 s22 s33 s58 s3 min0.78
Surreal-GAN91 s103 s2 min3 min5 min0.41
s-SuStaIn3 s36 s90 s4 min23 min1.33
BEBMS2 min8 min28 min100 min–1.78
SuStaIn (GMM)2 min14 min99 min9.8 h–2.70
SuStaIn (Z-score)32 min11.3 h––––
Dash indicates unavailable result due to its long fitting time.

We will add these results to the revised manuscript.

Reviewer_h2qFOpenReview ↗

reply

I thank the ahthor for the relies and raise the score to 3.

Thank you!

Reviewer_3eATOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TRACE is a probabilistic framework that jointly estimates disease subtypes and continuous stages from MRI biomarkers. It represents subtype-specific trajectories using monotonic neural networks and incorporates repeated observations through a change-reconstruction loss. TRACE achieves strong subtype and stage recovery on simulated data and identifies progression patterns associated with clinical differences in two representative benchmarks, ADNI and OASIS-3.

Critical strengths
  1. Jointly modeling subtype and stage is well motivated and provides a coherent way to account for progression when grouping bomarker profiles.
  2. TRACE shows strong subtype and stage recovery against relevant baselines in the main semi-simulated experiments. Ablations also support the usefulness of incorporating longitudinal changes.
Critical weaknesses

The certainty and longitudinal stability of subtype assignments could be characterized more clearly. Although TRACE estimates subtype probabilities and reports follow-up consistency, the analyses do not show how confidently individuals are assigned or whether subtype changes occur primarily in uncertain cases (e.g., 0.49vs0.51 and 0.1vs0.9 are both subtype1). This makes it difficult to assess whether assignment changes reflect ambiguous boundaries, measurement variability, or remaining dependence on progression stage.

Questions for authors
  1. Could you report the distribution of subtype posterior confidence and distinguish assignment changes near uncertain boundaries, such as 0.51 versus 0.49, from changes involving confident assignments?
  2. Given the shared-subtype assumption during training, are visit-level assignment changes associated with progression in inferred stage? Would joint inference across an individual’s visits improve consistency without compromising fit?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 3eAT

We thank the reviewer for thoughtful comments and constructive suggestions. In response, we conducted additional analysis to quantify confidence in subtype assignments, characterize subtype switching across visits, and evaluate joint subtype inference from multiple visits of the same participant. We present the results below. We've also added Appendix B.5 and B.6 in the revised manuscript for these new results.

1. Weakness and Q1: Confidence in subtype assignments

In response to this question, we quantified confidence in subtype assignments in the real-data cohort using the fitted model with K=4K=4. For each scan, assignment confidence was defined as the highest posterior probability across the four subtypes.

Subtype assignments showed high overall confidence (Appendix B.5, Figure 8(a)), with a median maximum posterior probability of 0.977 across all scans. The maximum posterior probability was at least 0.6 for 91.8% of scans, 0.80 for 77.7%, 0.90 for 66.2%, and 0.95 for 56.8%. Uncertainty was concentrated at early inferred stages (Figure 8(b)): the median maximum posterior probability was 0.82 for stages 0–0.2, compared with at least 0.99 for later stages. This pattern is consistent with the model assumption that all subtype trajectories originate from a common normal state and their proximity in early progression stages makes subtype distinctions less pronounced and assignments correspondingly less certain.

2. Q1 and Q2: Do subtype switches occur near assignment boundaries, and are they associated with inferred stage or disease progression?

Subtype switching was substantially more frequent for uncertain assignments and at the earlier disease stages. We analyzed pairs of consecutive visits from the same participant. At each visit, we assigned the subtype with the highest posterior probability and recorded a switch when the assignments differed between visits. The switch rate was 8.6% among pairs in which both visits had confident assignments (maximum posterior probability >0.9>0.9), compared with 43.1% among pairs with at least one uncertain assignment (maximum posterior probability <0.6<0.6). Overall, approximately three quarters of all switches involved at least one visit with a maximum posterior probability ≤0.9\leq 0.9. We also visualize the distribution of switches along the disease stages (Fig. 8 (c)). Most switches (76.4%) fall at early and intermediate stages (0.0–0.6), and fewer at late stages (18.0% at 0.6–0.8 and 5.6% at 0.8–1.0), which is consistent to Fig. 8(b) that these stages have more non-confident visits.

3. Q2: Would joint inference across an individual's visits improve consistency without compromising fit?

Yes, joint inference across an individual's visits improve consistency without compromising fit. Our framework can integrate inference results across a participant's visits without refitting the model. Assume that all visits share a subtype, that observations are conditionally independent given the subtype and visit-specific stages. For participant ii with TT visits, the joint subtype posterior is:

\propto \pi_k \prod_{t=1}^{T} \int_0^1 p(x_{it}\mid k,s) p(s\mid k) ds \propto \frac{\prod_{t=1}^{T}p(k\mid x_{it})}{\pi_k^{T-1}},$$ where $x_{it}$ denotes the observations at visit $t$, and $\pi_k$ is the prior probability of subtype $k$. Under our model's uniform stage prior, $p(s\mid k)=1$ for $s\in[0,1]$, each integral is the visit-level marginal likelihood already computed in Eq. 6. Thus, the joint posterior can be obtained directly from the visit-level subtype posteriors, with a correction to count the subtype prior only once. We evaluated this procedure using the saved model $\lambda=30$ as shown in the updated Table 1. The fitted model parameters were held fixed, and only inference results across visits of the same individuals were integrated. Joint inference increased ARI from **0.903 to 0.964** and JSSC from **0.934 to 0.976**. Stage $\rho$ remained essentially unchanged (approximately +0.001), while longitudinal stage $\rho$ increased slightly (approximately +0.010). These results suggest that combining visits improves subtype recovery and consistency while preserving staging accuracy. | Assignment rule | ARI | Stage $\rho$ | L-Stage $\rho$ | JSSC | |:--|:--|:--|:--|:--| | Visit-level (Table 1) | 0.903 $\pm$ 0.023 | 0.878 $\pm$ 0.035 | 0.799 $\pm$ 0.044 | 0.934 $\pm$ 0.010 | | Joint, all visits | **0.964 $\pm$ 0.016** | **0.879 $\pm$ 0.034** | **0.810 $\pm$ 0.042** | **0.976 $\pm$ 0.005** | We've added these results and the joint inference procedure in the revised manuscript. We appreciate the reviewer's insightful comments, which motivated these additional analyses and strengthened our evaluation.
Reviewer_3eATOpenReview ↗

Response to Authors

Thank you for your efforts, and my concerns are largely addressed. While i remain positive about the paper, some concerns remain.

Figure 8(c) reports the proportion of all switches occurring in each stage interval rather than the switching rate within that interval. Also, my original question concerned progression between visits.

Overall, i will keep my score to weak acceptance.

Thank you for clarifying Q2 and for maintaining the score! We agree that Figure 8(c) reports the overall proportion rather than the switching rate conditional on each interval and we'll revise accordingly.

Regarding the relationship between subtype switching and stage progression, we further examined our results. The inferred stage may change after subtype switching because similar biomarker patterns can correspond to different stages in different subtypes. We did not observe a consistent pattern of stage changes associated with subtype switching, as the inferred stage can either increase or decrease depending on which subtype the subject switches from and to. We consider this as an intrinsic ambiguity of performing inference independently at each visit, particularly when the fitted subtype trajectories overlap in terms of biomarker patterns. If longitudinal visits are available, using joint inference across visits as you suggested could help reduce this ambiguity and thus improve inference performance.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.