acceptodds
Under review as a conference paper at ICLR 2027

GraphTSA: Adaptive Token Communication for Local Heterogeneity and Global Synchronization in Medical Time Series

Abstract

Medical time series (MedTS) exhibit a structural duality: measurements retain local heterogeneity from distributed sources and measurement geometry, while sharing global synchronization from common physiological activity. Dense pairwise attention captures heterogeneous interactions but lacks explicit global coordination; CoTAR efficiently aggregates global information but broadcasts a shared core across tokens, limiting heterogeneous communication. We propose Graph Token Structural Attention (GraphTSA), an MLP-based module that preserves CoTAR's token-message construction but replaces shared broadcast with input-adaptive, directional structural mixing without a separate value projection. A low-dimensional query–key factorization controls affinity-logit capacity. Across seven public EEG, ECG, and human-activity-recognition datasets, GraphTSA improves upon TeCh with CoTAR on six. Rank and topology analyses reveal dataset-dependent rank sensitivity, richer heterogeneous communication, and a persistent dominant spectral mode. Despite replacing TeCh's linear token mixing with adaptive pairwise communication, GraphTSA incurs less than 1% parameter and memory overhead on PTB at its best-performing rank. Code and training scripts will be released upon acceptance.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Medical Time Series, Graph Learning, Low-Rank Attention, Interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_FDRQOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Graph Token Structural Attention (GraphTSA), a token-mixing module for multivariate medical time series. It retains CoTAR-style token-message construction but replaces shared-core broadcast with an input-dependent, directed affinity matrix generated by low-dimensional query and key projections. The method is inserted into TeCh’s temporal and channel branches and evaluated on five EEG/ECG and two HAR classification datasets.

Critical strengths
  • The paper is motivated by a meaningful and practically relevant observation. The dual structure of local heterogeneity and global synchronization reflects a genuine characteristic of multivariate medical time series, providing a reasonable and well-motivated foundation for the proposed method.
  • The method demonstrates a thoughtful balance between relational expressiveness and model complexity. Constructing input-dependent relational structures through low-dimensional query and key projections is a reasonable design choice, while the analyses of matrix rank, topology.
  • The experimental results demonstrate competitive performance against existing methods. The reported results suggest that the proposed approach is a promising alternative to several established baselines, supporting the potential value of modeling input-dependent relationships in multivariate time series.
Critical weaknesses

Weakness 1: The structural motivation relies on established characteristics of medical time series rather than a clearly differentiated methodological contribution.

  • Local temporal heterogeneity and cross-channel dependencies are well-recognized characteristics of multivariate medical time series and have motivated a substantial body of research on medical time-series modeling. For example, Medformer (NeurIPS 2024)[1] explicitly models cross-channel correlations and temporal patterns at multiple granularities, while MedMamba (ICML 2026)[2] explores local temporal morphology, multi-view representations, and adaptive graph learning for medical time-series classification. These studies suggest that leveraging local temporal patterns and cross-channel relationships is an established research direction rather than a novel observation introduced by GraphTSA. Therefore, the paper should more clearly distinguish the established structural properties of medical time series from its actual methodological contributions. In particular, the authors should explain what limitations remain in existing approaches that already exploit local temporal patterns and cross-channel dependencies, and demonstrate through controlled comparisons that GraphTSA's adaptive communication mechanism addresses these limitations in a distinct and effective manner. Otherwise, the claimed structural motivation risks overstating the novelty of the work, while the specific advantages of the proposed mechanism over existing alternatives remain insufficiently established.

Weakness 2: The contribution of the proposed communication mechanism is not sufficiently isolated.

  • Although GraphTSA outperforms the original TeCh with CoTAR in several settings, the experiments do not adequately establish that the gains specifically result from the proposed adaptive communication mechanism. GraphTSA is structurally close to a single-head scaled dot-product attention layer, yet the paper lacks a parameter-matched attention baseline within the same TeCh architecture and a controlled comparison using conventional value projections. Moreover, GraphTSA introduces an additional internal residual path, whose effect is not adequately controlled in the comparison with CoTAR. More targeted ablations with matched residual strengths and alternative token-mixing mechanisms are needed to isolate the contribution of the proposed design.

Weakness 3: The evidence for the proposed structural motivation and the interpretability of learned communication patterns is insufficiently representative.

  • The paper uses selected examples from a few datasets to visualize learned communication topologies and motivate the claim that GraphTSA captures local heterogeneity and global synchronization. However, such qualitative examples are vulnerable to sample-selection effects and do not establish that the observed patterns are representative or consistent across the data distribution. A systematic evaluation across all seven datasets and a substantially larger number of samples is needed to determine whether these communication patterns persist beyond a few illustrative cases. Furthermore, the analysis should include misclassified samples to examine whether the learned affinities remain meaningful when the model fails. Without such evidence, the visualizations provide limited support for the claimed structural motivation and do not establish that the learned communication patterns reflect generalizable properties of multivariate medical time series.

Weakness 4: The interpretation of affinity rank as dataset-dependent communication capacity is insufficiently supported.

  • The rank constraint (k) limits the bilinear affinity logits, but it does not directly determine the rank or effective complexity of the post-softmax communication matrix. The claimed relationship between rank sensitivity and data compressibility is based primarily on the contrast between PTB and FLAAP, while PCA compressibility does not directly establish the appropriate rank of learned affinity logits. In addition, the analysis does not disentangle the temporal and channel branches, despite their different token counts and semantics. A more rigorous analysis should clarify the validation-based selection of (k), examine the two branches separately, and evaluate the claimed association across all seven datasets.

Weakness 5: The practical advantage of GraphTSA over competitive CNN/TCN baselines is insufficiently established.

  • The paper primarily compares GraphTSA with existing Transformer-based methods, leaving its advantage over competitive convolutional architectures unclear. Notably, the simple HM-BiTCN+CIF combination has already demonstrated strong performance before the ICLR 2027 submission deadline, suggesting that effective medical time-series classification may not necessarily require the more elaborate token-communication mechanism proposed by GraphTSA. Without comparisons against strong CNN/TCN baselines under comparable experimental settings, it remains unclear whether GraphTSA provides meaningful improvements in predictive performance or offers a favorable trade-off between accuracy and computational cost. The authors should include competitive convolutional baselines, particularly HM-BiTCN+CIF [3], and provide a comprehensive comparison of predictive performance and computational efficiency to substantiate the practical necessity and advantages of the proposed architecture.

Reference [1] Medformer: A Multi-Granularity Patching Transformer for Medical Time-Series Classification [2] MedMamba: Multi-View State Space Models with Adaptive Graph Learning for Medical Time Series Classification [3] Channel-Imposed Fusion: A Simple yet Effective Method for Medical Time Series Classification

Questions for authors

Question 1: What is the methodological novelty of GraphTSA beyond established structural properties of medical time series?

Local temporal heterogeneity and cross-channel dependencies have already been extensively studied in medical time-series modeling, including Medformer (NeurIPS 2024) and MedMamba (ICML 2026). What specific limitations of these existing approaches does GraphTSA address? How does the proposed adaptive communication mechanism differ fundamentally from existing methods that already model local temporal patterns and inter-channel relationships? Can the authors provide controlled experiments demonstrating that the proposed mechanism offers a distinct advantage beyond these established design principles?

Question 2: Can the performance gains be specifically attributed to the proposed adaptive communication mechanism?

Given the structural similarity between GraphTSA and single-head scaled dot-product attention, can the authors provide parameter-matched attention baselines within the same TeCh architecture, including comparisons with conventional value projections? Furthermore, how much of the observed improvement comes from the additional internal residual path rather than the proposed communication mechanism itself? Can controlled ablations with matched residual strengths and alternative token-mixing mechanisms isolate the contribution of each design choice?

Question 3: Are the learned communication patterns representative and consistent across the data distribution?

The paper visualizes selected examples from a few datasets to support its claim that GraphTSA captures local heterogeneity and global synchronization. How do the authors establish that these patterns are representative rather than artifacts of sample selection? Can they systematically evaluate communication topologies across all seven datasets and a substantially larger number of samples? In addition, do the learned affinities remain meaningful for misclassified samples, and how do their patterns differ between correctly and incorrectly classified cases?

Question 4: Does the affinity rank reflect dataset-dependent communication capacity, and is this interpretation empirically justified?

The rank constraint kk limits the bilinear affinity logits but does not directly determine the rank or effective complexity of the post-softmax communication matrix. How do the authors justify the claimed relationship between affinity-rank sensitivity and data compressibility, given that PCA compressibility does not directly determine the appropriate rank of learned affinity logits? Can the authors clarify how kk is selected using the validation set, analyze the temporal and channel branches separately, and verify this relationship across all seven datasets?

Question 5: Does GraphTSA offer a meaningful practical advantage over competitive CNN/TCN baselines?

The evaluation primarily compares GraphTSA with Transformer-based methods, while the HM-BiTCN+CIF combination has already demonstrated strong performance on medical time-series classification benchmarks. Can the authors include HM-BiTCN+CIF and other competitive CNN/TCN baselines under comparable data splits and evaluation protocols? Does GraphTSA consistently improve predictive performance over these simpler architectures, and do any gains justify its additional architectural complexity and computational cost? Without such comparisons, how can the practical necessity and advantages of the proposed communication mechanism be established?

10 comments
AuthorsOpenReview ↗

Response to Reviewer FDRQ (Part 1)

Weakness 1 and Question 1: Methodological Novelty Beyond Established MedTS Properties

We thank the reviewer for recognizing the practical relevance of our structural motivation, the thoughtful balance between relational expressiveness and complexity, and the competitive empirical results. We agree that temporal heterogeneity, inter-channel dependencies, and adaptive graph learning are established research directions, including in Medformer and MedMamba. We do not claim novelty in these physiological properties or in query–key attention itself. Our contribution is a specific communication design: combining CoTAR-derived token messages with input-adaptive, receiver-specific aggregation within TeCh.

  • Distinction from prior methods. Medformer models multi-granularity temporal patterns and cross-channel correlations; MedMamba uses multi-view state-space representations and adaptive graph learning. GraphTSA instead examines a particular limitation of CoTAR's aggregation stage. CoTAR forms an input-dependent shared core and subsequently fuses it nonlinearly with each token; thus, it can produce heterogeneous final representations. However, the shared aggregate itself does not assign different source-token mixtures to different receivers. GraphTSA explicitly permits this:

    A(X)=\operatorname{softmax}_{\mathrm{row}}\!\left(\frac{Q(X)K(X)^\top}{\sqrt D}\right),$$ where $O_j(X)$ is a CoTAR-derived transformed and reweighted source message. The low-dimensional query–key projections produce directional, sample-conditioned source weights, while retaining CoTAR-style message construction rather than a separate conventional value projection. **This is a targeted architectural inductive bias, not a fundamentally new attention primitive.**
  • Controlled evidence for the design. We evaluated seven communication and residual configurations with the same TeCh backbone, fixed subject-independent PTB partition, and five seeds.

    Table 1. Selected PTB operator comparisons (mean ± SD, five seeds).

    ConfigurationAccuracy (%)F1 (%)
    Parameter-matched attention84.50 ± 2.0679.79 ± 3.14
    Fixed learned adjacency85.11 ± 2.4580.52 ± 3.96
    Normalized uniform aggregation87.23 ± 2.3783.79 ± 3.50
    GraphTSA (full)89.11 ± 1.1686.53 ± 1.64

    GraphTSA exceeds parameter-matched attention by 4.61 Accuracy and 6.74 F1 percentage points, and fixed adjacency by 4.0 and 6.01 points, respectively. Uniform aggregation is competitive (differences of 1.88 and 2.74 points), confirming the value of shared communication. These comparisons use a separate controlled PTB setting from the main benchmark and support the complete communication formulation rather than isolating every design factor. The full seven-configuration study includes conventional attention and internal-residual controls (Appendix K, Table 12).

  • Structural evidence beyond selected examples. The three-way decomposition in Section 4.4.3 and Figure 6(b) quantifies uniform, shared nonuniform, and receiver-specific affinity components. Appendix H, Figure 9 and Table 10 extend this analysis to all seven datasets, every enabled temporal/channel branch, and all test samples and graph layers. The observed mixtures vary by dataset, branch, and depth, showing that GraphTSA does not impose one universal communication pattern.

  • Contribution and manuscript clarification. Section 3.1 and Appendix K.1, Equation (14) clarify the relationship to CoTAR and conventional attention. Taken together, the controlled comparisons and broad structural analyses support GraphTSA as an empirically useful architectural extension of CoTAR. We interpret the structural analysis as descriptive and the ablations as evidence for the combined design, rather than proof of a uniquely causal mechanism.

AuthorsOpenReview ↗

Response to Reviewer FDRQ (Part 2)

Weakness 2 and Question 2: Isolating Adaptive Communication, Attention, and Residual Effects

We thank the reviewer for recognizing GraphTSA's balance between relational expressiveness and model complexity. We agree that the strongest test of this design is not simply outperforming TeCh, but determining whether its gains persist against parameter-matched attention, fixed/shared aggregation, and residual controls. We therefore conducted seven controlled ablations on PTB, directly targeting the proposed alternatives.

  • Controlled protocol. All seven configurations use the same TeCh backbone, fixed subject-independent PTB partition, three channel-token layers, validation-F1 checkpoint selection, and five seeds. GraphTSA uses k=8k=8 and 1/D1/\sqrt D logit scaling. Results below are mean ± SD; all six metrics and operator definitions appear in Appendix K.1–K.3, Table 12.

    Table 1. Controlled PTB operator and residual ablations.

    ConfigurationAccuracy (%)F1 (%)
    GraphTSA (full)89.11 ± 1.1686.53 ± 1.64
    Parameter-matched attention84.50 ± 2.0679.79 ± 3.14
    Conventional attention85.88 ± 4.3781.62 ± 6.55
    Fixed learned adjacency85.11 ± 2.4580.52 ± 3.96
    Normalized uniform aggregation87.23 ± 2.3783.79 ± 3.50
    GraphTSA without internal residual85.54 ± 2.9781.33 ± 4.66
    CoTAR-style residual control84.54 ± 2.1379.87 ± 3.41
  • Beyond conventional attention and parameter count. With the TeCh backbone held constant, GraphTSA exceeds parameter-matched QKV attention by +4.61 Accuracy and +6.74 F1 points. It also exceeds the conventional-attention replacement by +3.23 and +4.91 points. Both attention alternatives use learned value projections; GraphTSA instead routes CoTAR-derived transformed and reweighted messages through input-dependent affinities. This is direct evidence that the proposed message-construction-plus-routing formulation is useful beyond substituting a generic attention layer. Parameter matching concerns registered model parameters.

  • Why adaptive source selection matters. Holding CoTAR-derived messages in the comparison, replacing sample-dependent affinities with a fixed learned adjacency reduces Accuracy by 4.0 points and F1 by 6.01 points. Replacing them with normalized uniform aggregation reduces Accuracy by 1.88 and F1 by 2.74 points. The latter remains competitive and achieves the highest AUROC (95.27%) and AUPRC (95.08%) in this controlled study. This is an important result: shared communication is valuable, but adaptive, receiver-specific source selection yields the best mean Accuracy and F1 among all seven variants. The uniform control is a normalized mixing alternative, not a full CoTAR reproduction.

  • Quantifying the internal residual's contribution. Removing only GraphTSA's internal residual, while retaining its adaptive operator and the outer TeCh residual, decreases Accuracy by 3.57 points and F1 by 5.20 points. The CoTAR-style residual control reaches 84.54% Accuracy / 79.87% F1, further demonstrating that adding residual communication to shared aggregation does not reproduce the full GraphTSA result. The residual helps; it does not, by itself, explain the advantage over the tested attention and aggregation alternatives.

  • Conclusion and manuscript revisions. These targeted controls answer the reviewer's central concern: GraphTSA ranks first in mean Accuracy and F1 across all seven tested configurations, including parameter-matched attention and residual alternatives. The comparisons identify meaningful contributions from both adaptive communication design and residual placement; they do not require attributing every point of improvement to a single component. Appendix K.1–K.4, Table 12 provides the mathematical equivalence to attention (Equation (14)), full protocol, six-metric results, and interpretation.

AuthorsOpenReview ↗

Response to Reviewer FDRQ (Part 3)

Weakness 3 and Question 3: Representativeness of Learned Communication Patterns

We thank the reviewer for recognizing the relevance of our structural motivation and the value of the rank/topology analyses. We agree that selected affinity heatmaps alone are insufficient. To directly address this concern, we extended the analysis from illustrative examples to every test sample, every enabled temporal/channel branch, and every graph layer across all seven datasets, using five seeds.

  • Full-distribution evaluation. The analysis covers 10 dataset–branch combinations and 71,529 held-out test samples per seed, totaling 357,645 sample evaluations across five seeds. The same held-out partitions are evaluated across seeds. For each affinity matrix, we quantify the proportions of squared Frobenius energy assigned to three orthogonal components:

    A=U+(C−U)+(A−C),U=11⊤N,C=UA.A=U+(C-U)+(A-C),\quad U=\frac{\mathbf 1\mathbf 1^\top}{N},\quad C=UA.

    The components represent uniform aggregation, shared nonuniform source preferences, and receiver-specific deviations. We first aggregate across test samples and layers within each branch, then report mean ± SD across five seeds.

  • Evidence across datasets and branches. The results show that GraphTSA learns distinct communication regimes depending on the signal and tokenization, rather than producing a fixed pattern across datasets.

    Table 1. Selected full-test-set communication-energy fractions (%; mean ± SD across seeds).

    Dataset / BranchUniformShared nonuniformReceiver-specific
    ADFTD / Channel99.78 ± 0.130.20 ± 0.130.02 ± 0.00
    ADFTD / Temporal58.46 ± 7.9820.58 ± 3.6020.96 ± 5.03
    APAVA / Channel97.76 ± 0.741.95 ± 0.700.29 ± 0.05
    PTB / Channel72.44 ± 2.2217.28 ± 1.5010.27 ± 2.29
    PTB-XL / Temporal23.02 ± 2.8955.25 ± 3.9321.73 ± 1.93
    TDBrain / Temporal67.63 ± 2.5215.39 ± 1.9616.98 ± 0.86

    ADFTD provides a particularly informative within-dataset contrast: its channel branch is almost entirely uniform (99.78%), while its temporal branch retains 20.96% receiver-specific energy. On PTB-XL, shared nonuniform source preferences dominate the temporal branch (55.25%), alongside 21.73% receiver-specific energy. Thus, the learned operator adapts its balance of shared and heterogeneous communication to the dataset and branch. Appendix H, Table 10 provides all 10 combinations.

  • Layer-level evidence, not cherry-picked samples. We also analyze all layers and identify the layer with the highest five-seed mean receiver-specific energy in each branch. TDBrain's temporal branch reaches 38.48 ± 3.02% at layer 1/9, compared with 16.98 ± 0.86% averaged across its layers; PTB-XL reaches 33.22 ± 13.07% at layer 1/7. These values reveal substantial depth-dependent specialization. We explicitly distinguish the maximum-selected layers from branch-wide averages; Figure 9(a–b) presents both views.

  • Correct versus incorrect predictions. We include misclassified test samples rather than restricting analysis to successful predictions. The receiver-specific energy difference (correct minus incorrect, percentage points) is +1.46 ± 0.91 for PTB channel, +1.10 ± 0.65 for UCI-HAR channel, and +0.57 ± 0.46 for FLAAP temporal, but −0.36 ± 0.21 for ADFTD temporal and −0.90 ± 0.29 for TDBrain temporal. Therefore, heterogeneous communication remains measurable in the full test distribution, but greater receiver-specific energy is not a universal indicator of correct classification.

  • Conclusion and manuscript revisions. This full-distribution analysis directly addresses the reviewer's sample-selection concern: the shared/heterogeneous communication patterns are quantified across all seven datasets, all active branches and layers, and both correct and incorrect predictions. Their variability is itself informative: GraphTSA can favor nearly uniform aggregation where appropriate and differentiated communication elsewhere. Section 4.4.3, Equation (12) defines the decomposition; Appendix H, Figure 9 and Table 10 report the complete five-seed analysis. The affinities represent learned computational dependencies.

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

The paper proposes Graph Token Structural Attention (GraphTSA), motivated by the coexistence of local heterogeneity and global synchronization in medical time series. Built on the previous state-of-the-art CoTAR model, GraphTSA replaces its shared-core broadcast with input-adaptive, token-specific aggregation, using a low-dimensional query–key factorization to control the rank of the affinity logits. Experiments on five EEG/ECG datasets and two human activity recognition datasets report improvements over TeCh with CoTAR on six of the seven datasets.

Critical strengths
  1. Structural view of MedTS. The paper offers a useful perspective on medical time series by considering both local heterogeneity and global synchronization.

  2. Broad empirical evaluation. The experiments cover multiple datasets and a wide range of baselines, providing a broad assessment of the proposed method.

Critical weaknesses
  1. About the motivation: The introduction characterizes pairwise attention and shared aggregation as two communication extremes, suggesting that the latter may underrepresent heterogeneous token relationships. However, CoTAR already enables indirect communication between tokens through an input-dependent core token and combines the shared context with each token’s own representation.

In this regard, it would therefore be helpful to clarify which specific heterogeneous relationships are difficult to capture through this mechanism and why token-specific pairwise aggregation is needed.

  1. About the method: The novelty novelty in incremental. GraphTSA primarily extends CoTAR by introducing input-adaptive graph construction through a low-dimensional query–key factorization (Equations 2–4), while retaining its re-weighted token-message construction and TeCh’s dual-tokenization architecture. The methodological contribution therefore appears incremental, yet the motivation of this modificaiton needs more clarifications.

  2. About the experiment: In table 2 and 3, the improvements of GraphTSA over CoTAR appear modest compared with the gains for CoTAR over earlier baselines. Apart from PTB, most gains are small (relative to the reported variation across runs), and performance decreases on FLAAP.

Questions for authors
  1. Which heterogeneous relationships are difficult for CoTAR to capture, and can you demonstrate this through a controlled experiment? Why this kind of heterogeneous relationships is important?

  2. What experimental evidence directly links the performance gains to the joint modeling of local heterogeneity and global synchronization?

2 comments
AuthorsOpenReview ↗

Response to Reviewer iFFj (Part 1)

Weaknesses 1–2 and Question 1: Heterogeneous Relationships and Receiver-Specific Communication

We thank the reviewer for recognizing the value of our global-synchronization/local-heterogeneity perspective, broad evaluation, and methodological soundness. We agree that CoTAR's core is input-dependent and that its token-specific nonlinear fusion can produce heterogeneous outputs. Our claim is narrower: CoTAR's aggregation uses the same source mixture for every receiving token, whereas GraphTSA explicitly learns different source mixtures for different receivers and recordings.

  • What CoTAR's shared core does not explicitly represent. Let Oj(X)O_j(X) denote the CoTAR-derived reweighted message of source token jj. At the aggregation level,

    c(X)&=\sum_j O_j(X),\\ m_i(X)&=\sum_j A_{ij}(X)O_j(X), \end{aligned}$$ where $c(X)$ is CoTAR's shared aggregate and $m_i(X)$ is GraphTSA's receiver-specific aggregate. CoTAR subsequently combines $c(X)$ with each token through nonlinear fusion, so its final representations need not be identical. However, **the aggregate itself cannot assign one source-token weighting pattern to receiver $i$ and another to receiver $i'$**. GraphTSA does so through directional, input-conditioned $A_{ij}(X)$. We do not claim that CoTAR is theoretically incapable of approximating such dependencies through its full nonlinear layer.
  • Why this matters for MedTS. ECG leads reflect synchronized activity from the same heart but record distinct spatial projections. Consequently, the most useful information sources may differ by receiving lead and by recording. Shared aggregation captures coordinated information; receiver-specific aggregation permits different leads to emphasize different sources. These affinities are learned computational relationships.

  • Controlled experimental evidence. We compared seven communication and residual configurations using a common TeCh backbone, fixed subject-independent PTB partition, and five seeds. The table highlights five configurations (mean ± standard deviation).

    Table 1. Controlled PTB operator comparison.

    ConfigurationAccuracy (%)F1 (%)
    Normalized uniform aggregation87.23 ± 2.3783.79 ± 3.50
    Fixed learned adjacency85.11 ± 2.4580.52 ± 3.96
    Conventional attention85.88 ± 4.3781.62 ± 6.55
    GraphTSA without internal residual85.54 ± 2.9781.33 ± 4.66
    GraphTSA89.11 ± 1.1686.53 ± 1.64

    GraphTSA exceeds normalized uniform aggregation by 1.88 Accuracy and 2.74 F1 percentage points, and fixed adjacency by 4.00 and 6.01 points, respectively. This supports the practical value of input-conditioned communication relative to shared or fixed aggregation in this setting. Importantly, uniform aggregation is competitive, indicating that shared communication remains valuable. The residual ablation also shows that the full-model gain cannot be attributed solely to adaptive routing.

  • Methodological contribution and manuscript clarification. We acknowledge that low-dimensional query–key factorization is established attention machinery. The contribution is combining CoTAR-derived token messages with explicit, receiver-dependent aggregation within TeCh, rather than proposing a new attention primitive. Section 3.1 describes the module; Appendix K.1, Equation (14) clarifies its relationship to CoTAR and attention; Appendix K, Table 12 reports all seven operator and residual configurations.

AuthorsOpenReview ↗

Response to Reviewer iFFj (Part 2)

Weakness 3 and Question 2: Evidence for Joint Modeling of Global Synchronization and Local Heterogeneity

We thank the reviewer for asking how the proposed structural motivation relates to predictive improvements. We agree that higher classification accuracy alone does not demonstrate joint modeling. We therefore combine direct measurements of learned communication with controlled operator comparisons, while distinguishing structural observations from causal evidence.

  • Quantifying shared and receiver-specific communication. For each learned row-stochastic affinity matrix, we use the orthogonal decomposition

    A=U+(C−U)+(A−C),U=11⊤N,C=UA,A=U+(C-U)+(A-C),\qquad U=\frac{\mathbf{1}\mathbf{1}^{\top}}{N},\quad C=UA,

    where the terms represent uniform aggregation, shared nonuniform source preferences, and receiver-specific deviations. Their squared Frobenius norms partition the matrix's communication energy. On 9,368 PTB test segments across five seeds, the results are:

    Table 1. GraphTSA communication-energy fractions (%).

    LayerUniformShared nonuniformReceiver-specific
    Layer 258.526.315.2
    Layer 362.922.514.1

    Thus, substantial shared communication coexists with nonuniform source preferences and receiver-dependent deviations. Layer 1 is even more uniform (95.6%), suggesting that different depths learn different mixtures of shared and heterogeneous communication.

  • Spectral and null-baseline evidence. GraphTSA's effective ranks across three channel-token layers are 1.54/2.39/2.45, compared with 3.08/4.02/3.54 for conventional attention and 1 for CoTAR's normalized uniform aggregation reference. Independently sampled Dirichlet row-stochastic matrices and row-shuffled GraphTSA affinities have substantially higher effective ranks. The latter control preserves individual row weights but disrupts source assignments, indicating that GraphTSA's spectral concentration reflects coordinated source preferences rather than normalization or per-row distributions alone.

  • Link to predictive performance. In the controlled PTB operator study (same TeCh backbone and five seeds), full GraphTSA obtains 89.11 ± 1.16% Accuracy / 86.53 ± 1.64% F1, versus 87.23 ± 2.37% / 83.79 ± 3.50% for normalized uniform aggregation and 85.11 ± 2.45% / 80.52 ± 3.96% for fixed learned adjacency. These comparisons support the usefulness of the complete adaptive communication design relative to shared and fixed aggregation in this setting. As noted in Q1, the residual controls show that performance gains cannot be assigned exclusively to receiver-specific routing. The decomposition itself is not a component-removal experiment, so these observations do not prove that each communication component independently causes the improvement.

  • Magnitude and limitations of the gains. We acknowledge that gains over TeCh are modest on several datasets and that FLAAP is the exception: GraphTSA improves Accuracy and F1 on six of seven datasets but underperforms on FLAAP. PTB shows the largest improvement among the five medical datasets (+3.15 Accuracy and +4.56 F1 percentage points versus reproduced TeCh). Our PCA and rank-sensitivity analyses (Section 4.3.2, Figure 4; Appendix I.1, Table 11) further explain this contrast: PTB exhibits greater linear compressibility and stronger rank sensitivity, whereas FLAAP shows limited channel redundancy and nearly rank-insensitive performance, suggesting that the benefits of adaptive communication depend on the underlying signal structure. The paired tests on PTB and PTB-XL in Appendix D.2, Table 8 yield several p<0.05p<0.05 results. We therefore present the results as encouraging, dataset-dependent evidence, not conclusive superiority across tasks.

  • Manuscript revisions. The communication decomposition is reported in Section 4.4.3 and Figure 6(b); the spectral and null-baseline analyses in Section 4.4.2, Figure 6(c), and Appendix G, Table 9; and the controlled operator results in Appendix K, Table 12. The dataset-dependent performance and FLAAP exception are discussed in Section 4.2, with statistical limitations in Appendix D.2.

Reviewer_R5MpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GraphTSA, an adaptive token-communication module for medical time-series classification. GraphTSA preserves the token messages used by CoTAR while replacing its shared broadcast mechanism with an input-dependent directional affinity matrix constructed from low-dimensional query and key projections. The module is integrated into the temporal and channel branches of TeCh.

Critical strengths
  1. The paper provides a clear, controlled comparison against TeCh by replacing its CoTAR communication component while largely retaining the surrounding architecture.
  2. The evaluation covers multiple modalities and datasets, uses several random seeds, and adopts subject-independent splits for the medical datasets.
  3. The authors transparently report complexity, latency, memory, and the distinction between the rank of the logits and that of the softmax affinity matrix.
Critical weaknesses
  1. The core operation is close to single-head dot-product attention with low-dimensional queries and keys and a customized value representation. Moreover, under the 1/D1/\sqrt{D} scaling, changing kk changes both the rank capacity and the effective softmax temperature, so the rank conclusions are not fully isolated.
  2. Most improvements over TeCh are small, one dataset shows a regression, and no paired significance tests or alternative subject splits are reported. The spectral analysis lacks suitable null baselines, while the method remains O(N2)O(N^2) and increases latency by 22.422.4% on PTB.
Questions for authors
  1. How does GraphTSA differ technically from single-head attention using low-dimensional QQ and KK projections and CoTAR messages as values?
  2. How was kk selected for each dataset? Could the rank experiments be repeated using 1/k1/\sqrt{k} scaling or matched attention temperatures?
  3. Are the improvements statistically significant under paired tests or multiple subject-level splits?
  4. Could the spectral results be compared with standard attention, CoTAR, and untrained or random row-stochastic matrices?
4 comments
AuthorsOpenReview ↗

Response to Reviewer R5Mp (Part 1)

Weakness 1 and Question 1: Technical Distinction from Self-Attention and Methodological Novelty

We thank the reviewer for highlighting the relationship between GraphTSA and low-dimensional single-head attention. We agree that query–key factorization is not itself novel. GraphTSA's aggregation is algebraically equivalent to single-head attention using CoTAR-derived messages as values, under matching affinity definitions. Our contribution is therefore not a new attention primitive, but the structurally motivated integration of CoTAR-style message construction with input-adaptive, directional token communication within TeCh.

  • Physiological motivation. Multichannel medical time series (MedTS) capture shared physiological activity through heterogeneous measurements. For example, ECG leads observe electrical activity from the same heart through different spatial projections. Clinically, ST-segment changes in anatomically related leads and reciprocal changes in other leads contribute to myocardial infarction interpretation [1,2]. This motivates modeling both global synchronization and local heterogeneity. CoTAR emphasizes global coordination through a shared core, whereas GraphTSA allows each receiving token to aggregate a different combination of CoTAR-derived source messages.

  • Mathematical relationship. Let X∈RN×DX\in\mathbb{R}^{N\times D} denote input tokens and O(X)O(X) denote CoTAR-derived reweighted messages. At the aggregation level, the mechanisms can be expressed as

    M_{\mathrm{uniform}} &= A_{\mathrm{uni}}O(X),\\ M_{\mathrm{GraphTSA}} &= A(X)O(X),\\ M_{\mathrm{attention}} &= A(X)V(X), \end{aligned}$$ where $A_{\mathrm{uni}}=\mathbf{1}\mathbf{1}^{\top}/N$, $A(X)=\operatorname{softmax}_{\mathrm{row}}(QK^\top/\sqrt{D})$, and $V(X)$ denotes conventional learned value representations. **If $V(X)=O(X)$, the attention and GraphTSA aggregation expressions coincide.** The architectural distinction from conventional attention lies in retaining CoTAR's nonlinear, token-reweighted messages rather than introducing a separate learned value projection. The distinction from CoTAR lies in replacing shared aggregation with input-conditioned, receiver-specific source weighting. CoTAR can nevertheless produce token-dependent outputs through subsequent nonlinear fusion. These equations compare aggregation mechanisms, not complete layers or their residual structures.
  • Controlled experimental comparison. In the main PTB benchmark, GraphTSA achieves 89.11% Accuracy and 86.53% F1 (five-seed means). To examine the communication design, we additionally evaluated seven operator and residual configurations under a common TeCh backbone, fixed subject-independent PTB partition, and five seeds.

    Table 1. Controlled PTB operator comparison (mean ± standard deviation, five seeds).

    OperatorAccuracy (%)F1 (%)
    Conventional attention85.88 ± 4.3781.62 ± 6.55
    Parameter-matched attention84.50 ± 2.0679.79 ± 3.14
    Fixed learned adjacency85.11 ± 2.4580.52 ± 3.96
    Normalized uniform aggregation87.23 ± 2.3783.79 ± 3.50
    GraphTSA89.11 ± 1.1686.53 ± 1.64

    GraphTSA achieves the highest mean Accuracy and F1 among all seven configurations, exceeding conventional attention by 3.23 and 4.91 percentage points, respectively, and parameter-matched attention by 4.61 and 6.74 points. The competitive uniform-aggregation results demonstrate the value of shared information, while improvements over conventional attention and fixed adjacency support the effectiveness of combining CoTAR-derived messages with adaptive aggregation. The complete ablation also includes residual controls, distinguishing the communication design from the contribution of the additional internal residual.

  • Manuscript revisions. We clarify the attention-style formulation and its relationship to CoTAR in Appendix K.1, Equation (14). The complete seven-configuration comparison, including residual controls, appears in Appendix K, Table 12. Complementary structural evidence is provided in Section 4.4 and Figure 6.

References

[1] Goldberger et al. Goldberger's Clinical Electrocardiography: A Simplified Approach. Elsevier.

[2] Thygesen et al. Fourth Universal Definition of Myocardial Infarction (2018). European Heart Journal, 2019.

AuthorsOpenReview ↗

Response to Reviewer R5Mp (Part 2)

Weakness 1 and Question 2: Rank Selection and Rank–Temperature Coupling

We thank the reviewer for identifying the interaction between projection rank and attention scaling. We agree that changing kk can affect both query–key representation capacity and softmax sharpness. We therefore clarify our rank-selection procedure and report an additional scaling comparison.

  • Rank selection and mathematical clarification. GraphTSA computes

    Ak(X)=softmax⁡row(Qk(X)Kk(X)⊤D),A_k(X)=\operatorname{softmax}_{\mathrm{row}}\left(\frac{Q_k(X)K_k(X)^\top}{\sqrt{D}}\right),

    where DD is the token embedding dimension and kk is the query–key projection dimension. The factorization bounds the rank of the pre-softmax logits by kk; it does not generally bound the rank of the resulting affinity matrix. We select kk separately for each dataset using validation performance, without test-set selection. PCA is an independent analysis of input structure, not a preprocessing step or rank-selection rule. On PTB, the original rank sweep peaks at k=8k=8, while nine of 15 principal components explain 95% of the variance (Section 4.3, Figure 4).

  • Additional attention-scaling experiment. To examine the reviewer's proposed alternative, we compared the original 1/D1/\sqrt{D} scaling with 1/k1/\sqrt{k} on PTB. Both sweeps use k∈{2,4,6,8,10,12,14,16}k\in\{2,4,6,8,10,12,14,16\}, learning rate 3×10−53\times10^{-5}, and five seeds (42–46). Figure 10 in Appendix I.2 reports the full Accuracy and F1 curves with standard-deviation error bars.

    Table 1. Peak PTB results under two scaling conventions (five-seed means).

    ScalingPeak rank kkAccuracy (%)F1 (%)
    Original 1/D1/\sqrt{D}889.1186.53
    Alternative 1/k1/\sqrt{k}687.9985.03

    The preferred rank shifts from 8 to 6, confirming that rank-dependent performance is sensitive to the scaling convention. Nevertheless, both sweeps favor intermediate ranks rather than the largest evaluated rank, consistent with the usefulness of controlling the capacity of token-interaction logits. The original scaling also achieves higher peak mean Accuracy and F1 in this PTB comparison; we do not interpret this as proof that it is universally preferable.

  • Interpretation and scope. Standard scaled dot-product attention normalizes by the square root of the query–key dimension [1]. GraphTSA instead uses D\sqrt{D} in its original formulation. Because the learned query–key magnitudes can vary with kk and training, neither normalization ensures matched effective softmax temperatures across independently trained models. Thus, the additional experiment tests sensitivity to the scaling convention but does not fully separate rank capacity from attention temperature. The rank/PCA association is consequently presented as descriptive evidence rather than a mechanistic identification of the optimal kk.

  • Manuscript revisions. Appendix I.2, Figure 10 contains the two scaling sweeps and their interpretation. Appendix I.1, Table 11 provides the dataset-level rank/PCA comparison, and Section 4.3, Figure 4 retains the original PTB/FLAAP rank-sensitivity analysis.

Reference

[1] Vaswani et al. Attention Is All You Need. NeurIPS, 2017.

AuthorsOpenReview ↗

Response to Reviewer R5Mp (Part 3)

Weakness 2 and Question 3: Statistical Significance and Robustness Across Subject-Level Splits

We thank the reviewer for raising the question of statistical significance and robustness. We conducted two-sided paired t-tests comparing GraphTSA and TeCh (CoTAR) across five corresponding training seeds on the established subject-independent PTB and PTB-XL partitions.

  • Paired statistical comparisons. The tests cover six classification metrics on each dataset (12 comparisons in total). Table 1 reports the representative Accuracy and F1 results. Improvements are GraphTSA minus TeCh in percentage points; the reported pp-values are unadjusted.

    Table 1. Seed-paired comparisons (five seeds).

    DatasetMetricImprovement (pp)Unadjusted pp
    PTBAccuracy+4.580.0161
    PTBF1+6.820.0163
    PTB-XLAccuracy+0.390.0211
    PTB-XLF1+0.430.1022

    PTB Accuracy and F1 and PTB-XL Accuracy reach the conventional unadjusted p<0.05p<0.05 threshold, whereas PTB-XL F1 does not. Given the 12 metric–dataset comparisons and the limited number of seeds, we interpret these as suggestive evidence of improved mean performance.

  • Baseline provenance and comparability. For these paired tests, TeCh results come from the original authors' publicly released training logs, with each method evaluated using its respective reported configuration. They are not the reproduced TeCh results in the main Table 2. Accordingly, the paired-test improvements above differ from the main controlled PTB benchmark gains of +3.15 Accuracy and +4.56 F1 percentage points. We report both comparisons transparently rather than combining results from different baseline sources.

  • Subject-level robustness. Our medical-dataset evaluation uses subject-independent partitions, preventing subject overlap between training and testing. However, the five seeds correspond to repeated training on fixed partitions, not five independently sampled subject splits. The paired tests therefore quantify variability across training runs under those partitions.

  • Manuscript revisions. The test protocol, unadjusted pp-values, baseline provenance, and split-related limitations are documented in Appendix D.2, Table 8, with a corresponding reference in Section 4.2. These analyses complement the controlled TeCh comparison and five-seed mean ± standard-deviation results in Table 2.

AuthorsOpenReview ↗

Response to Reviewer R5Mp (Part 4)

Weakness 2 and Question 4: Spectral Analysis and Null Baselines

We thank the reviewer for recognizing the controlled TeCh comparison, multi-modal evaluation, subject-independent splits, and transparent efficiency reporting. To address the remaining spectral concern, we extended the analysis to conventional attention, CoTAR's normalized uniform aggregation reference, and two stochastic controls. Results cover all three channel-token layers, 9,368 subject-independent PTB test segments, and five training seeds.

  • Spectral comparison. We measure the effective rank of post-softmax affinity matrices, using the entropy of their normalized singular-value spectra.

    Table 1. PTB effective rank (mean ± SD across five seeds).

    OperatorLayer 1Layer 2Layer 3
    CoTAR uniform reference1.001.001.00
    Conventional attention3.08 ± 0.594.02 ± 0.873.54 ± 0.53
    GraphTSA1.54 ± 0.182.39 ± 0.362.45 ± 0.20
    Row-shuffled GraphTSA4.05 ± 1.358.96 ± 0.518.87 ± 0.32
    Dirichlet row-stochastic10.35 ± 0.0010.35 ± 0.0010.35 ± 0.00

    GraphTSA lies between rank-one uniform aggregation and the more distributed spectra of conventional attention. Figure 6(c) shows a dominant singular mode with nonzero secondary components. The CoTAR entry is an analytical normalized uniform reference, not the full nonlinear CoTAR layer.

  • Null-baseline validation. Independent Dirichlet rows test whether row-stochastic normalization alone produces the observed concentration. Row-shuffling preserves each GraphTSA receiver row's affinity weights and sum while randomizing source assignments. Both controls have substantially higher effective rank. Thus, GraphTSA's concentration is not explained by row normalization or per-row weight distributions alone; coordination of source preferences across receivers contributes to the learned structure.

  • Global and heterogeneous communication. We additionally decompose AA into orthogonal uniform, shared nonuniform, and receiver-specific components (Section 4.4.3, Equation (12)). Their squared-Frobenius-energy shares are 58.5%/26.3%/15.2% in Layer 2 and 62.9%/22.5%/14.1% in Layer 3. This directly supports our architectural motivation: GraphTSA retains substantial shared communication while learning differentiated source preferences and receiver-specific interactions. These affinities describe learned computational dependencies.

  • Computational complexity and measured efficiency. We also address the reviewer's concern about quadratic complexity and increased latency. GraphTSA explicitly constructs an N×NN\times N affinity matrix; its additional communication cost is O(NDk+N2k+N2dc)\mathcal{O}(NDk+N^2k+N^2d_c) for NN tokens, embedding dimension DD, projection rank kk, and message dimension dcd_c. The pairwise terms remain quadratic in NN. On PTB, the channel branch uses 15 tokens and the temporal branch is disabled for optimal performance.

    Table 2. PTB performance and inference efficiency.

    MethodAccuracy (%)Parameters (M)Latency (ms)Peak memory (MB)
    TeCh (CoTAR)85.961.5490.902 ± 0.01162.69
    GraphTSA (k=8k=8)89.111.5611.104 ± 0.01462.74

    On one NVIDIA H100 NVL (batch size 128), GraphTSA gains 3.15 Accuracy points with 0.80% more parameters, 0.08% more peak memory, and 22.4% higher latency. We regard this as a measurable speed trade-off, not negligible overhead. Across k∈{1,4,8,16}k\in\{1,4,8,16\}, latency ranges from 1.104 to 1.169 ms.

  • Manuscript revisions. Section 4.4, Figure 6, and Appendix G, Table 9 document the spectral and null comparisons; Section 4.5, Figure 7, and Appendix L, Tables 13–14 report complexity, profiling protocol, and rank-dependent efficiency.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Reviewers and AC,

We sincerely thank the reviewers for their time, thoughtful feedback, and continued engagement throughout the discussion. We have revised the manuscript to incorporate the additional experiments, analyses, and methodological clarifications presented in our rebuttals. To facilitate review, newly added sections and subsections are identified by red headings, while substantially revised sections and subsections are identified by blue headings. The main updates are summarized below.

  1. Clarification of methodological novelty and structural motivation. We refined the positioning of GraphTSA relative to CoTAR, conventional attention, and existing MedTS methods. We clarify that our contribution lies in combining CoTAR-derived token messages with input-adaptive, receiver-specific communication, rather than introducing a fundamentally new attention mechanism or discovering previously unknown physiological properties.

  2. Additional controlled communication and residual ablations. We evaluated seven communication configurations on PTB, including parameter-matched attention, conventional attention, fixed learned adjacency, normalized uniform aggregation, and residual controls. GraphTSA achieves the highest mean Accuracy and F1 among these configurations. We also clarify the contributions of adaptive communication and residual design, while acknowledging the PTB-specific scope of these experiments.

  3. Expanded cross-dataset structural analysis. We extended the communication analysis to all seven datasets and 71,529 held-out test samples per seed, using five training seeds. The results quantify uniform, shared nonuniform, and receiver-specific communication across different datasets, branches, and layers, demonstrating that GraphTSA learns dataset-dependent communication patterns rather than imposing a fixed structure.

  4. Additional rank, spectral, and training-dynamics analyses. We expanded the graph-rank investigation to all seven datasets, compared PCA-derived dimensionality with validation-optimal graph rank, examined alternative attention-scaling conventions, and introduced spectral and null-baseline analyses. We also investigated the evolution of learned communication during training. We explicitly distinguish these empirical observations from theoretical or causal conclusions.

  5. Expanded baseline comparisons and computational analysis. We added CNN/TCN comparisons across seven datasets, further examined GraphTSA's computational efficiency relative to TeCh, and clarified the distinction between parameter efficiency and practical runtime performance.

  6. Additional statistical analyses and clearer limitations. We incorporated paired statistical comparisons, expanded experimental details, and clarified limitations regarding dataset-dependent performance, the scope of controlled ablations, and the interpretation of learned communication structures.

We are grateful for the opportunity to strengthen our manuscript through this constructive discussion. Throughout the revision, we have sought to provide additional experimental evidence, clarify our methodological contributions, and acknowledge the limitations of our findings.

We particularly appreciate the reviewers' recognition of the importance of capturing both global synchronization and local heterogeneity in medical time series. While these properties have been explored in prior research, we emphasize that GraphTSA's methodological contribution lies in translating this structural motivation into an adaptive, receiver-specific communication mechanism that extends CoTAR's shared aggregation. We believe the additional experiments and analyses provide stronger empirical support for this design and further clarify its distinctions from existing approaches.

We hope that the revised manuscript offers a clearer understanding of GraphTSA's motivation, methodological novelty, and empirical contributions, and that the reviewers will consider these improvements in their final assessment.

Sincerely,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.