acceptodds
Under review as a conference paper at ICLR 2027

Context-Conditioned Tokenization for Medical Time Series Classification

Abstract

Local discriminative patterns in medical time series often depend on neighboring waveforms and the overall physiological state, whereas conventional token formation compresses local observations before contextual modeling, potentially weakening fine-grained cues that can only be identified with sufficient context. To address this issue, we propose ReConTok, a context-conditioned token formation framework for medical time series classification. Its core module, Context-guided Token Gating and Fusion (CoTGF), jointly exploits the current local representation, neighborhood context, and global summary to conditionally gate and fuse fine-grained features. The same mechanism is applied to both temporal and channel tokens while leaving the subsequent Transformer interaction unchanged. Across five EEG/ECG datasets, ReConTok achieves the best F1 on four datasets and shows consistent advantages in Recall. Controlled tokenizer comparisons, short-versus-long window experiments, and robustness analyses further validate its effectiveness. These results demonstrate that directly incorporating context into token formation yields more discriminative and robust representations for medical time series. blueCode is available in the supplementary material.

Primary area:
learning on time series and dynamical systems
Keywords:
Medical Time Series Classification, Context-Conditioned Token Formation, Tokenization

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_q5ahOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a model that exploits the pattern characteristics of medical time series. Before tokens enter the Transformer, it combines information from the current time point, its neighborhood, and the global context at the token formation stage. This reflects a key property of medical time series well, namely that the meaning of a local pattern depends on its surrounding context and the overall physiological state. I also appreciate that the paper goes beyond comparing classification performance. It analyzes the model from several angles, including ablations, robustness tests, and patient-level analysis.

However, a few aspects need improvement. First, the mathematical formulation should be organized more clearly, with all symbols defined and a consistent flow between equations. Second, the experiments mainly demonstrate gains in performance and robustness. They do not directly verify whether the proposed structure works as intended, for example whether the gates actually assign higher weights to meaningful positions depending on the context. If these points are addressed, I believe this would be a solid contribution.

Critical strengths
  1. The paper takes a well-motivated direction by jointly considering three key aspects of medical time-series patterns: the local observation itself, its neighboring context, and the global state of the sample. I find this a sound way to approach the problem.

  2. The proposed model is analyzed from multiple perspectives, including ablations, a controlled tokenizer comparison, robustness tests, and patient-level analysis. These experiments are well suited to demonstrating the merits of the architecture.

  3. The evaluation protocol is careful. All experiments use subject-independent splits, which avoids data leakage across subjects. The paired short- and long-window setup on PTB-XL keeps labels, sample counts, and splits fixed, so it isolates the effect of the available temporal context.

Critical weaknesses
  1. The proposed idea is well motivated, but it is difficult to verify whether the architecture actually implements the intended mechanism. Some symbols are used without definition, and the flow between equations is occasionally broken. The method section needs a clearer and more self-contained formulation, ideally with a notation table.

  2. To show that the proposed design is meaningful, an interpretability analysis is needed. For example, the authors could visualize the learned gates GiG_i and mixing coefficients αiα_i and check whether the weights concentrate on specific ECG leads or on clinically relevant segments (e.g., the QRS complex or ST segment). Without such evidence, it is hard to tell whether the gains come from context-conditioned token formation as claimed.

Questions for authors

[Questions]

  1. What is the specific reason for choosing the SRS tokenizer as the controlled comparison?
  2. Several symbols in Eqs. (1)–(4), such as AA, AggAgg, PoolPool, and N(i)N(i), are not clearly defined. Please define each symbol when it is first introduced.
  3. It remains unclear whether the proposed structure actually realizes the described methodology. In Eq. (7), the gate multiplies the input features before a nonlinear MLP and a second softmax inside PbP_b, so a larger gate value does not necessarily mean a larger contribution to the aggregation. Could the authors explain how the design guarantees the intended behavior, or provide empirical evidence (e.g., the learned values of ρρ and αα)?
  4. Are 1 s and 10 s appropriate definitions of "short" and "long"? Clinically, 10 s still seems short. I understand this is a dataset limitation, but what is the practical difference in diagnostic content between the two settings?
  5. In Section 5.4, a performance drop under noise is expected. Wouldn't the relative degradation (e.g., F1_noisy / F1_clean) be more informative than absolute F1, especially since ReConTok starts from the highest clean score?
  6. Table 9 uses different branch configurations per dataset (e.g., Channel only for PTB, Temporal only for PTB-XL). Why is this, given that applying CoTGF to both branches is claimed as a key contribution? Were the same configurations used for Table 2?

[Optional]

  1. Transposing Table 4 (swapping rows and columns) would likely improve readability.
  2. Figure 2(a) is too small, and its legend and bubbles are hard to read. Please consider enlarging it.
Reviewer_ytTMOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ReConTok, a framework for medical time-series classification that incorporates sample-conditioned processing into token formation. Its core module, CoTGF, constructs temporal and channel tokens by combining token reference representations with neighborhood information and global context. It then adjusts token features through context-conditioned fine-grained gating and global information fusion, followed by encoding using a standard Transformer encoder. Experiments on five EEG/ECG datasets report classification results that are competitive with some existing methods.

Critical strengths
  • The paper addresses a reasonable and practically relevant modeling problem: local EEG/ECG morphologies may only become meaningful in the context of surrounding signals and cross-channel structures, yet token formation is often treated as a fixed front end. Placing sample-conditioned computation before token interaction is a reasonable design direction, rather than simply proposing another variant of attention mechanisms.

  • The paper reports consistently competitive results across multiple datasets compared with Medformer and TeCh.

Critical weaknesses

Weakness 1: The central claim is not adequately supported by the ablation studies.

  • The ablations do not isolate the paper’s central claim, namely that current, neighborhood, and global context during token formation are responsible for the reported gains. In particular, the authors do not provide separate ablation studies for neighborhood aggregation and global aggregation, making it difficult to determine whether these components are necessary or whether the improvements primarily arise from other aspects of the proposed module. More concerningly, in the released implementation (model.py, line 128), neighborhood = torch.zeros_like(reference) initializes the neighborhood representation to zeros, and I could not identify any subsequent operation that incorporates actual neighborhood information into this variable. If this implementation reflects the computation used in the reported experiments, it raises a fundamental concern about whether the proposed neighborhood aggregation is actually implemented as described. The authors should clarify this discrepancy and provide sufficient implementation details and controlled experiments to substantiate their claims. Furthermore, since the primary contribution is a token representation mechanism rather than a fundamentally new Transformer architecture, it would be important to evaluate the proposed module across different Transformer variants. Such experiments would help establish whether the claimed benefits are attributable to the proposed token formation strategy itself, rather than to its particular integration with the current backbone.

Weakness 2: Insufficient reporting of dataset-specific hyperparameters undermines reproducibility and comparison fairness.

  • The paper does not provide the complete dataset-specific hyperparameter configurations used in the experiments. For example, Fig. 2(a) compares the inference speed of different models on ADFTD, yet Appendix C states that “We report the search ranges rather than dataset-specific final training settings.” Reporting only the search ranges, without disclosing the final configurations selected for each dataset, makes it difficult for readers to reproduce the results or verify whether the comparisons are conducted under fair and consistent settings. I strongly encourage the authors to provide the complete experimental configurations for every dataset, including the final hyperparameter values and relevant training and model-selection settings. In addition, a table summarizing the hyperparameter search process and the corresponding results would substantially improve the transparency, reproducibility, and credibility of the experimental evaluation.

Weakness 3: The baseline selection is insufficiently comprehensive, particularly regarding CNN-based methods for medical time-series classification.

  • Most of the baselines considered in this paper are Transformer-based models. However, convolutional neural networks (CNNs) have also demonstrated strong performance in time-series classification, including medical applications, and should be considered to provide a more comprehensive evaluation. Relevant methods include SGN (NeurIPS), FIC-TSC (ICML), InceptionTime, and Omni-Scale CNNs [1–4]. Their inclusion or, at a minimum, a more substantive discussion would help clarify the relative merits of the proposed approach beyond comparisons with Transformer-based architectures.

  • More importantly, I note that HM-BiTCN + CIF, a method based on a relatively simple temporal convolutional network (TCN) combined with Channel-Imposed Fusion, reports results that outperform those presented in this paper on the relevant datasets [5]. This work was publicly available well before the ICLR 2027 submission deadline and therefore represents a relevant reference point for evaluating the proposed method. The authors should discuss this comparison and, where experimental settings and evaluation protocols are compatible, include direct empirical comparisons. Without such comparisons, it remains unclear whether the proposed token formation strategy offers meaningful advantages over simpler convolution-based alternatives.

References

[1] SGN: Shifted Window-Based Hierarchical Variable Grouping for Multivariate Time Series Classification. NeurIPS.

[2] FIC-TSC: Learning Time Series Classification with Fisher Information Constraint. ICML.

[3] InceptionTime: Finding AlexNet for Time Series Classification.

[4] Omni-Scale CNNs: A Simple and Effective Kernel Size Configuration for Time Series Classification. ICLR.

[5] Channel-Imposed Fusion: A Simple yet Effective Method for Medical Time Series Classification.

Questions for authors

Question 1: Does the proposed neighborhood aggregation actually contribute to the reported performance gains?

  • Given that the ablation studies do not separately evaluate neighborhood and global aggregation, and the released implementation initializes neighborhood as a zero tensor, how can the authors demonstrate that neighborhood information is actually incorporated into token formation and contributes to the reported improvements? Could the authors provide controlled ablation experiments that independently remove neighborhood and global aggregation, as well as evaluations across different Transformer backbones, to verify the effectiveness and generalizability of the proposed mechanism?

Question 2: How can the experimental results be reproduced and their fairness verified without complete dataset-specific hyperparameter configurations?

  • Since Appendix C reports hyperparameter search ranges rather than the final configurations selected for each dataset, how can readers reproduce the reported results and verify that the comparisons are conducted under consistent and fair experimental settings? Could the authors release the complete dataset-specific configurations, including the final hyperparameter values, training settings, and model-selection criteria, together with a summary of the hyperparameter search process and its corresponding results?

Question 3: Does the proposed method offer meaningful advantages over simpler CNN-based alternatives?

  • Given the limited coverage of CNN-based baselines and the reported results of HM-BiTCN + CIF, which outperforms the proposed method on relevant datasets, what evidence supports the superiority of the proposed token formation strategy over simpler convolution-based approaches? Could the authors include direct comparisons with representative CNN-based methods, particularly HM-BiTCN + CIF, under compatible experimental settings and evaluation protocols, to establish whether the proposed method provides consistent and meaningful performance improvements?
Reviewer_7Ap4OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for context-conditioned token construction tailored for medical time-series classification. The core module integrates local representations, neighborhood information, and global summaries during the token-formation stage. The authors evaluate the proposed framework on five EEG and ECG datasets and provide additional analyses, including ablation studies, experiments with varying window lengths, and robustness evaluations.

Critical strengths
  1. The paper focuses on information loss during the tokenization stage, and thats a meaningful idea.

  2. The intuition of using the current token, neighborhood information, and global state to jointly modulate local features is sound.

  3. The experiments cover five EEG/ECG datasets and explicitly employ subject-independent data partitioning.

  4. The paper is well-structured.

Critical weaknesses
  1. The technical novelty is limited. CoTGF primarily combines existing methods, such as dynamic feature aggregation, neighborhood context, global pooling, and gated fusion. The paper does not state the fundamental technical differences between the proposal and prior works.

  2. The paper lacks fair, controlled comparisons in the main experiments. Except for TDBRAIN, most baseline results are taken from existing literature. The observed performance gains may reflect differences in implementation details.

  3. Some performance improvements are marginal, and the paper reports only the mean and standard deviation across five training seeds using a fixed subject split. These metrics capture only the randomness of the optimization process.

  4. The paper calculates the robustness auc using the absolute F1 under perturbations, rather than the relative performance degradation compared to clean inputs. More experiments are helpful for strengthening the experiments.

  5. In some cases, certain metrics for the ablated models are even lower than those obtained using a linear tokenizer. Also, the paper does not perform ablation studies by removing the current, neighborhood, and global contexts.

Questions for authors

Please refer to weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.