acceptodds
Under review as a conference paper at ICLR 2027

Learning Shared Visual Exploration Primitives for Cross-Subject Mental Disorder Diagnosis from Eye Movements

Abstract

Cross-subject mental disorder diagnosis from eye movements requires learning disease-discriminative representations that generalize beyond the subjects observed during training. This is challenging because eye-movement behaviors exhibit substantial inter-subject variability, such that the same disorder may manifest through heterogeneous visual exploration strategies. Existing methods may collapse individual-specific pathological manifestations into a common representation, limiting generalization to unseen subjects. Motivated by shared structural patterns in eye movements across individuals, we propose to represent eye movements using shared, fine-grained, latent visual exploration primitives and their personalized compositions. Thus, we develop a shared Visual Exploration Primitive-Aware (VEPA) model for cross-subject mental disorder diagnosis. By jointly modeling shared primitives and personalized compositions, VEPA learns robust, disease-discriminative representations that generalize to unseen subjects. Specifically, our model first employs a Visual Primitive Discovery module (VPD) to model diverse visual exploration primitives as Gaussian distributions, whose correspondence with eye-movement patterns is characterized through post-hoc visualization. Then a Personalized Primitive Composition module (PPC) is designed to efficiently encode primitive compositions via differentiable Gumbel-Softmax routing and adaptive layer selection, with contrastive learning boosting classification. Experiments on two clinical eye-tracking datasets for depression and schizophrenia diagnosis demonstrate that our model outperforms existing methods, achieving a 2.3% relative accuracy improvement.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Cross-Subject, Mental Disorder Diagnosis, Eye Movements, Shared Visual Exploration Primitives

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_cr7GOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a shared Visual Exploration Primitive-Aware (VEPA) model for diagnosing cross-subject mental disorders. The author(s) claim that VEPA learns robust, disease-discriminative representations that generalize to unseen subjects. The proposed approach has two modules in the methodology: a Visual Primitive Discovery module (VPD) to model diverse visual exploration primitives as Gaussian distributions, whose correspondence with eye-movement patterns is characterized through post-hoc visualization, and a Personalized Primitive Composition module (PPC) which is designed to efficiently encode primitive compositions via differentiable Gumbel-Softmax routing and adaptive layer selection, with contrastive learning boosting classification. The paper conducts experiments on two datasets and reports a 2.3% relative accuracy improvement over existing methods.

Critical strengths
  1. The paper attempts to solve an important and challenging problem: cross-subject mental disorder diagnosis from eye movements. Its key strength is the attempt to solve it with an interesting perspective: shared visual exploration patterns from subject-specific compositions through two complementary modules. This is potentially important because eye-movement behavior varies across individuals, making it difficult to learn representations that generalize to unseen subjects. Addressing this challenge could make eye-tracking-based diagnosis more robust and clinically useful.
Critical weaknesses
  1. The experimental evaluation appears limited, with results reported on only two datasets and most of the analysis concentrated on a single dataset (ETDS). This raises concerns about the robustness and generalizability of the proposed method across datasets and clinical populations. Furthermore, the paper lacks sufficient analysis to explain important empirical observations, such as why VEPA often achieves lower recall. Without a more thorough investigation of these results, it is difficult to determine the method's practical reliability and understand its limitations.

  2. The paper does not clearly establish what fundamental insight or methodological advance distinguishes VEPA from existing approaches. Although the proposed modules are described, their individual contributions and the specific benefits of modeling shared visual exploration primitives are not sufficiently clear from the presentation. Additionally, some parts of the paper, such as Section 4.5 and Figure 4, appear insufficiently connected to the central research, making the overall narrative and the significance of the findings difficult to follow. These issues weaken confidence in the paper's novelty and make it challenging to assess the significance of its contribution.

Questions for authors
  1. Can the author(s) emphasize the novelty and significance of the paper clearly?
  2. What distinguishes VEPA from existing approaches?
  3. Can the method be generalized beyond the two reported datasets, possibly to a bigger dataset?
  4. Can the author(s) explain why VEPA often achieves lower recall?
  5. Why is the method compared against naive approaches such as SVM, LR, and RF?
  6. The author(s) did some hyperparameter sensitivity analyses. Can they explain the outcomes of the hyperparameter sensitivity analyses clearly?
  7. Can the author(s) explain why Section 4.5 and Figure 4 were added, and what the intended outcome of this section is?
Reviewer_gJ8EOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes VEPA for cross-subject mental disorder diagnosis from eye movements. The main idea is that existing cross-subject methods often try to remove subject-specific variability, while some of this variability could actually correspond to different ways the same disorder is expressed across subjects. VEPA therefore learns a set of shared latent eye-movement primitives and then builds a personalized combination of these primitives for each sample. The model contains a Visual Primitive Discovery module and a Personalized Primitive Composition module, followed by contrastive learning and classification. The method is evaluated on depression and schizophrenia datasets with subject-independent cross-validation.

I find the idea interesting and the paper is well presented with a reasonably broad experimental evaluation. However, I am not fully convinced that the experiments clearly isolate the benefit of the proposed personalized primitive composition, and the improvement over the strongest baselines is relatively small compared with the reported variability.

Critical strengths
  • Interesting motivation and formulation of inter-subject variability. I think the main idea is interesting: instead of treating all subject-specific differences as nuisance information, the paper tries to preserve part of this variability through personalized combinations of shared primitives. This is meaningful for clinical problems where the same disorder may appear differently across subjects.

  • Good experimental coverage. The authors evaluate the method on two datasets and compare it with several eye-tracking and cross-subject methods. They also test VEPA with different temporal backbones and obtain improvements across most of them. The ablation study also shows that removing the main components generally decreases performance.

Critical weaknesses
  • The comparison with related representation-learning methods is limited. The main contribution of VEPA is based on learning a dictionary of shared primitives and composing them differently for each subject. However, the experiments mainly compare against eye-tracking and cross-subject generalization methods. The paper does not compare with broader dictionary learning or representation learning approaches that are conceptually close to the proposed method. I think this makes it harder to evaluate how novel the primitive-composition idea really is and whether its advantage comes from the specific VEPA formulation or from using a learned compositional representation in general.

  • The improvement over the strongest baselines is relatively small compared with the reported variability. On ETDS, VEPA improves accuracy from 86.31% to 88.31% compared with CL-CS, while the standard deviations are quite large. On EMS, VEPA improves from 83.13% to 84.38% compared with MSNet, but the MSNet result is directly taken from the original paper and has no standard deviation. I think that statistical significance or paired fold-level comparisons would be important here before concluding that VEPA clearly outperforms the strongest baselines.

Questions for authors
  • Could the authors provide statistical significance tests or paired fold-level comparisons between VEPA and the strongest baselines MSNET, especially CL-CS on ETDS?
  • How were M, C, K, and the loss weights selected? Were these hyperparameters selected only using the training/validation subjects, or were the reported test-fold results used to choose them?
  • For ETDS, the paper mentions that the Q=16 samples per subject are generated using overlapping sliding windows of M=30 trajectories, but the amount of overlap or stride is not specified. Could the authors clarify this?
  • Can the authors provide a comparison with a simpler or matched-capacity baseline to better isolate the benefit of the personalized primitive composition itself from the additional architecture and losses?
  • The adaptive layer is selected only according to the single dominant primitive. Could the authors justify this choice?
  • Since the main contribution relies on learning a shared primitive dictionary and personalized compositions, could the authors clarify the relation of VEPA to broader dictionary-learning, codebook/prototype-based, or compositional representation-learning methods?
Reviewer_4HwcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces VEPA, a framework for cross-subject mental disorder classification using eye-tracking data. The authors argue that conventional subject-invariant representation learning may remove useful individual variations. Instead, they propose learning a shared dictionary of latent visual exploration primitives and allowing each subject's eye-movement patterns to be represented through different combinations of these primitives.

The framework consists of two modules. Visual Primitive Discovery (VPD) models primitives as Gaussian distributions, uses orthogonality regularization to encourage diversity, and applies Wasserstein distance for distribution matching. Personalized Primitive Composition (PPC) uses differentiable top-K routing, subject-specific regularization, adaptive layer selection, and contrastive learning to construct classification representations.

Critical strengths
  1. Relevant problem with a useful representation-learning angle. Cross-subject generalization is a key challenge in physiological signal analysis. Preserving useful individual variation, rather than removing it, is a reasonable direction. Modeling eye movements through shared latent components with input-dependent selection offers an alternative to strictly subject-invariant representations.

  2. Broad evaluation. The framework is tested on two clinical eye-tracking datasets against traditional and deep-learning baselines. Backbone comparisons, ablations, and hyperparameter studies provide preliminary evidence of performance improvements across different temporal encoders.

  3. Potential interpretability and practical applicability. The visualizations of learned primitives provide some insight into recognizable eye-movement patterns, such as refixation and focal viewing. The model also generalizes to unseen subjects without requiring subject-specific fine-tuning, which is useful for cross-subject classification.

Critical weaknesses
  1. Limited novelty; mechanism not isolated. The method mainly combines established components. Prototype-based cross-subject learning has been explored by Cai and Pan (2023), while sparse expert routing is well established (Shazeer et al., 2017). More closely related to gaze modeling, Liu and Lu (2026) already employ prototype-guided routing and specialized experts for unseen-subject generalization, albeit for a different task. Equation (5) composes primitives through an unweighted sum of selected vectors, making its distinction from sparse prototype selection unclear. Adaptive layer selection also resembles conditional expert routing. Without parameter-matched comparisons against prototype-routing or sparse MoE baselines, the gains cannot be confidently attributed to the proposed compositional principle rather than additional model capacity.

  2. Weak presentation and insufficient methodological clarity. The writing is repetitive and relies heavily on claims about shared primitives, personalized compositions, and generalization without sufficiently explaining their technical significance. The manuscript discloses AI assistance in methodology, implementation, and writing, which is not inherently problematic. However, several design choices remain insufficiently justified, making it difficult to distinguish the methodological contribution from established techniques.

  3. Numerical inconsistency in reported results. In Table 1, PCDG's ETDS precision (76.20%) and recall (91.90%) imply an F1 upper bound of approximately 83.3%, not the reported 86.38%, assuming consistent binary metric definitions. Averaging F1 across folds cannot explain this discrepancy, since the mean of per-fold F1 cannot exceed the F1 calculated from averaged precision and recall. This raises concerns about metric-reporting consistency. The authors should clarify the calculation and provide fold-level results.

  4. Insufficiently controlled baseline comparisons. Table 1 uses MSNet results taken directly from Song et al. (2025) without clearly establishing identical subject partitions, preprocessing, and model-selection procedures. This matters given VEPA's small improvement of 1.25 percentage points over MSNet on EMS. Furthermore, the original EMS benchmark reserves 48 subjects for independent testing and uses the remaining 160 for four-fold cross-validation (Song et al., 2025). The manuscript reports only cross-validation results. A matched MSNet comparison and evaluation on the reserved test set would strengthen the claims.

Questions for authors

Q1. How does VEPA differ from existing prototype-based methods such as PCDG and sparse mixture-of-experts approaches? Can the authors provide a parameter-matched comparison to demonstrate its novelty?

Q2. In Table 1, PCDG reports precision of 76.20%, recall of 91.90%, and F1 of 86.38%. Under standard binary metric definitions, the F1 should not exceed approximately 83.3%. Can the authors explain this discrepancy and provide the fold-level results?

Q3. Why were MSNet's results taken from the original paper rather than reproduced under the same experimental conditions? Can the authors confirm that the evaluation protocols match and report results on the separate 48-subject EMS test set?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.