acceptodds
Under review as a conference paper at ICLR 2027

Channel-Imposed Fusion: A Simple yet Effective Method for Medical Time Series Classification

Abstract

**A simple model with good features beats a complex model with poor features.** Deep learning in healthcare has increasingly embraced a model-centric paradigm, and medical time series is no exception, with increasingly complex architectures designed to capture richer temporal, cross-channel, and multi-scale dependencies. We argue that this paradigm may overlook a more fundamental bottleneck: the quality and discriminability of the input signals themselves. We therefore shift the focus from designing increasingly complex classifiers toward improving data representations before classification. To this end, we propose CIF, a simple, plug-and-play, and model-agnostic framework for enhancing medical time series representations. CIF can be seamlessly integrated with existing classification architectures without modifying their model designs. Extensive experiments across diverse medical time series datasets and representative classification architectures demonstrate that CIF consistently and substantially improves classification performance, providing strong evidence for the effectiveness of a data-centric alternative. **We hope this work encourages the community to reconsider the core of medical time series classification: should progress be driven primarily by data-centric strategies, model-centric design, or both?**

Primary area:
learning on time series and dynamical systems
Keywords:
Medical Time Series Classification, Channel Imposed Fusion (CIF), Data-centric

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_QxZ9OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Channel-Imposed Fusion (CIF), a lightweight input transformation for medical time series classification. CIF linearly combines selected channel pairs using two shared coefficients, with the aim of incorporating physiological or spatial relationships into the input representation. The coefficients can be fixed or learned jointly with a downstream classifier. The authors also introduce HMTCN, a temporal convolutional backbone with bidirectional feature mixing. The evaluation shows that CIF improves HMTCN’s mean accuracy on all five medical datasets, with gains of 3.81–8.23 percentage points on four datasets and 0.81 percentage points on PTB-XL.

Critical strengths
  1. Comprehensive experimental evaluation. The paper evaluates CIF across multiple medical time series datasets and backbone architectures, supported by comparisons with spatial preprocessing methods and ablation studies. This breadth provides substantial empirical support for the method’s effectiveness and compatibility with different classification models.

  2. Clear organization and presentation. The paper follows a coherent progression from motivation and method design to experimental evaluation and further analysis. The illustrations and extensive supplementary material help readers understand the proposed approach and examine the evidence supporting its conclusions.

Critical weaknesses
  1. Limited methodological novelty. CIF is essentially a constrained linear spatial transformation, closely related to established channel mixing and spatial filtering techniques. While its simplicity is practically appealing, the paper does not sufficiently establish what methodological advance it offers beyond a particular parameterization of these existing approaches.

  2. An overly restrictive fusion mechanism. CIF applies the same two coefficients to all selected channel pairs, imposing a common interaction pattern despite potentially heterogeneous physiological relationships. This prevents different pairs from independently learning cooperative or differential combinations and limits the flexibility of the proposed representation. The paper does not sufficiently justify why this shared transformation is appropriate across the selected channels.

Questions for authors

Why are two shared coefficients across all selected channel pairs sufficient?

3 comments
AuthorsOpenReview ↗

Novelty and Rationale for Shared Fusion Coefficients

We sincerely thank you for your professional and insightful comments. We hope our response adequately addresses your concerns.

Q1: Methodological Novelty and Design Motivation of CIF

The motivation of CIF is not to introduce another linear transformation, but to explore whether a structurally constrained and parameter-efficient interaction mechanism can effectively facilitate cross-channel information fusion in multivariate time-series representation learning without relying on complex model designs.

First, regarding methodological design, conventional channel mixing approaches typically rely on flexible channel-wise projections, learnable mixing matrices, or attention mechanisms. In contrast, CIF employs two shared coefficients to establish a unified interaction pattern across selected channel pairs, introducing an explicit structural inductive bias for cross-channel representation learning rather than relying on highly flexible channel mappings.

Second, regarding design motivation, physiological time-series data often suffer from limited training samples, heterogeneous signal quality, and low signal-to-noise ratios. Highly flexible channel transformations may increase optimization and configuration complexity and make learned interactions more sensitive to individual datasets. Assigning independent coefficients to each channel pair would increase the number of adjustable parameters from 2 to 2n2n, substantially expanding the configuration search space. Therefore, sharing two coefficients represents a deliberate trade-off between interaction flexibility, parameter efficiency, and configuration complexity.

Finally, regarding applicability, we additionally investigate CCA-based channel fusion as an alternative strategy and analyze the physiological structures underlying channel interactions in Appendix J. Furthermore, we evaluate CIF on non-medical time-series tasks in Section 4.2(4) to examine its applicability beyond physiological signal analysis.

CIF explores whether a simple, unified, and structurally constrained interaction principle can achieve effective cross-channel information fusion while controlling parameter and configuration complexity. We have further clarified its design motivation, relationship to existing channel mixing approaches, and scope of applicability in the revised manuscript.

Q2: Rationale for Shared Fusion Coefficients

We sincerely thank the reviewer for raising this important question.

First, we acknowledge that physiological relationships may vary across different channel pairs. Assigning pair-specific parameters could theoretically provide greater flexibility in capturing such differences.

Second, parameter sharing offers a reasonable balance between interaction modeling capacity and configuration complexity. CIF employs two shared coefficients to model interactions across all selected channel pairs, keeping the number of independently adjustable parameters fixed regardless of the number of pairs. Compared with pair-specific parameterization, this design avoids introducing addition al degrees of freedom, controls configuration complexity, and reduces potential sensitivity to limited training data and dataset-specific variations.

Third, from the perspective of physiological motivation and generalizability, CIF is inspired by physiological priors and aims to facilitate cross-channel information fusion by leveraging potential physiological relationships among channels. Rather than explicitly parameterizing each pairwise relationship, CIF translates this physiological motivation into a unified and lightweight interaction mechanism through shared coefficients. This design accommodates varying channel configurations without increasing the number of adjustable parameters as more channel pairs are selected. To further investigate the role of physiological characteristics and the applicability of this design, we explore CCA-based fusion as an alternative that exploits inter-channel correlations and examine the physiological relationships underlying channel interactions in Appendix J. Moreover, we evaluate CIF on non-medical time-series tasks in Section 4.2(4) to assess its applicability beyond physiological signal analysis.

Overall, CIF is not intended to explicitly model the full diversity of physiological relationships between individual channel pairs. Instead, it draws inspiration from physiological priors while accounting for practical considerations, including parameter efficiency, configuration complexity, and applicability across different data domains. This design retains the guidance of physiological priors for cross-channel information fusion while providing a simple, unified, and lightweight interaction mechanism.

Thank you for authors effort and reply. I think authors reply sorted most of my concern. I decide to raise my score

AuthorsOpenReview ↗

Thank you!

We sincerely appreciate your recognition and hope that our work can bring new insights to the community. Thank you!

Reviewer_cdEgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Channel-Imposed Fusion, a data-centric preprocessing step for multivariate medical time series classification. Given C channels, it fuses the first n with the last n through two shared scalars and passes the result to a small temporal backbone. A two-channel SNR argument fixes the signs of those scalars, same sign when signal correlation exceeds noise correlation and opposite signs otherwise. Evaluation covers five datasets, three of them EEG, against model baselines and spatial filters under subject-independent splits. With the same backbone the step gains up to eight accuracy points on most datasets and under one point on PTB-XL.

Critical strengths
  1. Breadth of the Controlled Comparison: Each main table fixes one backbone and varies only the spatial preprocessing across eight alternatives, so the comparison isolates the step rather than the architecture. CIF leads most of those cells and every F1 entry under HMTCN, so the ranking does not hinge on one metric.
  2. Transfer to Other Backbones: The appendix applies the same transformation to the remaining baseline architectures, and most of them improve. Crossformer on APAVA gains about ten points in accuracy, and MTST on TDBRAIN by nearly as much, so the effect is not specific to one backbone.
  3. Low Cost: The step adds two scalars per dataset and linear work per time step, so the overhead stays small. Figure 5 reports lower compute and a far shorter training time than the slowest baseline on APAVA.
Critical weaknesses
  1. The Mechanism Does Not Reach the Classification Claim: Appendix A's algebra is correct, but it scores only the fused channel, while the classifier also sees the channels the fusion leaves in place. On the replacement reading the map is block triangular, with determinant ana^n in one direction and bnb^n in the other, and the replaced-block coefficient is nonzero at the reported settings. The operator is therefore an invertible linear reparameterisation whose inverse a channel-mixing layer can absorb, which leaves the Bayes rule unchanged, so a per-pair SNR story cannot carry a classification gain. One per-dataset pair of scalars also serves every pair in that dataset, and the learned signs reverse both physiological stories the Introduction tells.
  2. Hyperparameters Are Not Separated From Evaluation: The grid searches in the appendix report the same metric as the main tables, and for ADFTD and PTB the winning rows reproduce the reported numbers exactly. APAVA is the exception that matters, since its highest-accuracy row is 86.54 while the paper reports 86.30 and calls that the best result. No section states that the search used the validation split, and Appendix N describes the settings as adjusted by empirical experience. The baselines receive no comparable per-dataset search, so the reported margin is not a like-for-like comparison.
  3. A Higher-Scoring Variant Is Kept Out of the Main Text: Table 18 reports a CCA variant at 87.91 accuracy on APAVA, against 86.30 for the configuration the paper reports. Appendix J calls CCA an exploratory method and states that its results are deliberately left out of the main text. That is a wide margin on the paper's lead dataset, and the stated reason for excluding it is scope, not performance. If another fusion rule does better under the same protocol, the contribution is a family of rules rather than one, and the reported configuration is not its strongest instance.
  4. The Operator Description Conflicts With the Reported Settings: Section 3.1 says the fused channels are concatenated with the originals, while Algorithm 1, Appendix B.2 and the equations overwrite one block. The two readings imply different input widths, and at every reported setting the two blocks overlap, so some replaced channels are also the partners of other pairs. No untouched middle channel therefore survives, contrary to the appendix's account of channels that the transformation leaves alone. The appendix's prose gives APAVA the backward branch where caption and winning row give the forward one, and the opposite sign for one ADFTD coefficient.
Questions for authors
  1. Since the map is invertible, what experiment separates a representation gain from a reparameterisation effect?
  2. Were the fusion settings chosen on the validation split, and can the APAVA grid winner be reconciled with the reported configuration?
  3. Why is the CCA variant held back, and how does it score on the other four datasets?
  4. Move the negative-transfer cases, such as Nonformer on PTB, into the main text and state when the method fails.
4 comments

We sincerely thank you for your professional and insightful comments. We hope our response adequately addresses your concerns. Q1

To address this concern, we conduct an additional controlled experiment by constructing a non-invertible variant of CIF, termed HMTCN + CIF + Zero. The key idea is to deliberately break the invertibility of the original fusion mapping while retaining its core fusion operation, and then examine whether the performance gains of CIF can still be preserved.

Mathematical formulation.

Let the input feature consist of two channel groups, x,y∈Rn\mathbf{x},\mathbf{y}\in\mathbb{R}^{n}. The original CIF performs bidirectional channel fusion, which can be expressed as

\begin{equation}\begin{bmatrix}\mathbf{z}_{1}\\\ \mathbf{z}_{2}\end{bmatrix} = \underbrace{ \begin{bmatrix}a\mathbf{I}_{n}&b\mathbf{I}_{n}\\\ \mathbf{0}&\mathbf{I}_{n}\end{bmatrix} }_{\mathbf{W}_{\mathrm{CIF}}}\begin{bmatrix}\mathbf{x}\\\ \mathbf{y}\end{bmatrix} \end{equation}

where aa and bb are shared fusion coefficients and In\mathbf{I}_n denotes the nn-dimensional identity matrix. The determinant of this transformation is

=(a)^n$$ Therefore, when $a\neq 0$, the original fusion mapping is invertible and theoretically preserves all input information. To explicitly break this property, we introduce a diagonal masking matrix $$\begin{equation}\mathbf{M}= \operatorname{diag}(1,\ldots,1,0),\end{equation}$$ which sets the last output channel to zero after the original fusion operation. The resulting transformation becomes $$\begin{aligned} \mathbf{z}_{\mathrm{zero}} &= \mathbf{M}\mathbf{W}_{\mathrm{CIF}} \begin{bmatrix} \mathbf{x}\\\ \mathbf{y} \end{bmatrix}\\ &= \underbrace{ \mathbf{M}\mathbf{W}_{\mathrm{CIF}} }_{\mathbf{W}_{\mathrm{zero}}} \begin{bmatrix} \mathbf{x}\\\ \mathbf{y} \end{bmatrix}. \end{aligned}$$ Since $\det(\mathbf{M})=0$, we have $$\begin{equation}\det(\mathbf{W}_{\mathrm{zero}}) = \det(\mathbf{M})\det(\mathbf{W}_{\mathrm{CIF}}) =0.\end{equation}$$ Thus, the modified transformation is strictly non-invertible. Importantly, this modification retains the original fusion coefficients and fusion operation, while explicitly discarding one output channel. **Experimental results and analysis.** We compare three configurations: (1) the original HMTCN baseline, (2) HMTCN + CIF with the original invertible fusion mapping, and (3) HMTCN + CIF + Zero with the non-invertible mapping. All configurations are evaluated under the same subject-independent protocol. The results reveal that the non-invertible variant consistently outperforms the original HMTCN baseline on all four datasets with available Zero results. Specifically, on APAVA, HMTCN + CIF + Zero achieves 89.08\% accuracy and 88.69\% F1, outperforming the original CIF by 2.78 and 2.98 percentage points, respectively. On ADFTD, PTB, and TDBrain, the non-invertible variant exhibits some performance degradation compared with the original CIF. Nevertheless, it still improves accuracy by 4.98, 2.51, and 7.20 percentage points, respectively, over the original HMTCN baseline. The corresponding F1 improvements are 3.82, 4.04, and 7.34 percentage points. These results provide empirical evidence that the effectiveness of CIF does not solely depend on preserving a fully invertible mapping. In particular, the persistence of performance gains after deliberately removing one output channel suggests that the benefits of CIF cannot be explained merely by an invertible reparameterization. | Datasets | Models | Accuracy | Precision | Recall | F1 score | AUROC | AUPRC | |:---:|:---|---:|---:|---:|---:|---:|---:| | **APAVA** | HMTCN | 82.49±1.40 | 82.38±1.79 | 81.20±1.32 | 81.60±1.39 | 91.10±1.63 | 91.30±1.71 | | | HMTCN + CIF | 86.54±1.55 | 86.33±1.88 | 86.10±1.11 | 86.08±1.46 | 94.43±0.70 | 94.57±0.67 | | | HMTCN + CIF + Zero | **89.08±2.76** | **89.57±2.73** | **88.69±1.58** | **88.69±2.58** | **96.89±0.60** | **96.92±0.52** | | **ADFTD** | HMTCN | 52.05±2.22 | 50.45±3.00 | 50.40±2.55 | 49.48±2.70 | 69.43±2.84 | 50.99±3.15 | | | HMTCN + CIF | **58.56±0.93** | **55.65±0.81** | **55.86±0.79** | **55.42±0.82** | **76.07±0.59** | **59.75±0.67** | | | HMTCN + CIF + Zero | 57.03±0.73 | 53.58±1.50 | 54.00±1.40 | 53.30±1.60 | 74.27±1.17 | 57.37±1.41 | | **PTB** | HMTCN | 81.87±1.87 | 86.50±1.24 | 73.49±2.90 | 75.84±3.20 | 94.20±0.29 | 93.04±0.45 | | | HMTCN + CIF | **88.29±1.45** | **90.66±1.48** | **83.21±2.02** | **85.59±1.96** | **94.28±0.93** | **93.78±1.11** | | | HMTCN + CIF + Zero | 84.38±2.08 | 88.11±1.86 | 77.29±2.96 | 79.88±3.01 | 93.69±1.57 | 93.00±1.97 | | **TDBrain** | HMTCN | 84.90±2.60 | 86.02±2.00 | 84.90±2.60 | 84.76±2.74 | 93.94±1.92 | 94.20±1.85 | | | HMTCN + CIF | **93.13±1.41** | **93.33±1.37** | **93.13±1.41** | **93.12±1.42** | **98.62±0.66** | **98.68±0.63** | | | HMTCN + CIF + Zero | 92.10±3.93 | 92.16±3.86 | 92.10±3.93 | 92.10±3.93 | 97.81±1.74 | 97.89±1.66 |

Q2:

We thank the reviewer for pointing out this issue. We strictly follow the experimental protocol of Medformer to ensure a fair and consistent evaluation. Specifically, the training, validation, and test sets are split at the subject level, ensuring that no subjects overlap across the three subsets. Model selection is performed exclusively on the validation set, where the checkpoint achieving the best validation performance is selected for final evaluation on the test set. All relevant experimental settings and implementation details are fully consistent with the original Medformer implementation.

Regarding the APAVA grid search, we identified that the best result was inadvertently omitted during our result compilation. To resolve this discrepancy, we repeated the corresponding experiments and verified that 86.54 is the best result under the evaluated fusion configurations. We have corrected the reported results and updated the corresponding content in the revised manuscript accordingly. We apologize for this oversight and thank the reviewer for helping us identify it.

Q3:

DatasetsModelsAccuracyPrecisionRecallF1 scoreAUROCAUPRC
APAVAHMTCN82.49±1.4082.38±1.7981.20±1.3281.60±1.3991.10±1.6391.30±1.71
HMTCN + CIF(CCA)87.91±1.6089.81±0.9385.85±2.1286.93±1.9196.97±0.5296.91±0.49
ADFTDHMTCN52.05±2.2250.45±3.0050.40±2.5549.48±2.7069.43±2.8450.99±3.15
HMTCN + CIF(CCA)53.31±1.4049.90±1.5550.21±1.5749.75±1.6370.08±1.2351.96±1.92
PTBHMTCN81.87±1.8786.50±1.2473.49±2.9075.84±3.2094.20±0.2993.04±0.45
HMTCN + CIF(CCA)84.49±4.8488.07±2.7877.59±7.6079.63±7.2494.27±2.3293.45±2.44
TDBrainHMTCN84.90±2.6086.02±2.0084.90±2.6084.76±2.7493.94±1.9294.20±1.85
HMTCN + CIF(CCA)91.48±2.5391.67±2.5291.48±2.5391.47±2.5397.47±1.5397.56±1.47
PTB-XLHMTCN72.92±0.8864.21±1.5660.60±0.7261.49±0.8289.71±0.7266.44±1.49
HMTCN + CIF(CCA)73.08±0.4464.10±0.9960.57±0.9361.62±0.6089.92±0.4666.71±0.73

Q4:

The performance degradation observed when directly applying CIF to Nonformer is primarily associated with hyperparameter transfer across different backbone architectures. Specifically, the CIF configurations were initially optimized for HMTCN and directly transferred to Nonformer without architecture-specific adjustment. Since different backbones employ different feature representations and interaction mechanisms, their optimal fusion configurations may not be identical. Consequently, directly transferring the HMTCN-optimized configuration can result in suboptimal feature fusion and negative transfer.

To investigate this issue, we further optimized the CIF hyperparameters specifically for Nonformer. The results are presented in Table below. As shown, directly transferring the HMTCN-optimized configuration (Nonformer + CIF (HMTCN)) leads to substantial performance degradation across all datasets. In contrast, using the architecture-specific configuration (Nonformer + CIF (New)) consistently improves performance over the original Nonformer. For example, on PTB, the accuracy increases from 78.93% to 80.46%, while the F1-score improves from 82.48% to 84.22%. Similar improvements are observed across the remaining datasets and evaluation metrics.

These results suggest that the observed negative transfer is primarily associated with applying backbone-specific fusion hyperparameters without adaptation, rather than an inherent incompatibility between CIF and Nonformer. They further highlight that the effectiveness of CIF depends on appropriately configuring its fusion strategy to match the feature characteristics and interaction mechanisms of the target backbone.

Following the reviewer's suggestion, we have moved these negative-transfer results and the corresponding analysis into the main text. We also explicitly discuss this limitation, clarifying that directly reusing CIF configurations optimized for one backbone may lead to performance degradation when transferred to another architecture, and that backbone-specific hyperparameter selection is necessary for reliable performance.

Table: Negative-transfer analysis of CIF on Nonformer under the subject-independent setup.

ModelsAccuracyPrecisionRecallF1 scoreAUROCAUPRC
Nonformer78.93±1.4682.48±1.5369.68±2.1571.50±2.5690.54±0.5987.78±1.46
Nonformer + CIF(HMTCN)71.89±3.8171.80±4.5869.44±3.5669.74±3.8470.55±2.9670.78±4.08
Nonformer + CIF(New)80.46±2.2084.22±2.4471.79±3.1073.89±3.5092.74±0.9790.30±1.26

Thank you for the revision. Reporting the APAVA row with its coefficients makes the grid in Table 26 readable, the split protocol is now stated in Section 4, and the negative transfer result is in the main text instead of being routed around. What I still cannot settle is the mechanism, and that is holding me back from raising the score.

What is settled

  • Coefficients are reported per dataset, so the reported row can be checked against the grid (Tables 3 and 26).
  • The split protocol follows Medformer (Section 4).
  • CCA covers all five datasets (Table 20).
  • The negative transfer passage states the drop (Section 4.2).

What needs action

  1. The mechanism. The fusion is a full-rank linear map, and a channel-mixing layer downstream can absorb it, so the reported gain has no source in the operation itself. The control that would settle this is a fixed random invertible mixing, drawn once and never tuned, evaluated at the reported settings. Because such a matrix carries far more parameters than the two scalars the method uses, please also say how you would match the comparison. This is the one experiment I would need.

  2. The CIF+Zero variant. It is not in the paper, and as defined it removes an input channel on three datasets, so the gain is as consistent with channel deletion as with non-invertibility. Please put the table in the paper and add a channel-preserving variant that removes rank without removing information.

  3. Where the corners were run. The corner experiments use a different width from the reported rows, so they are compared against numbers produced under another setting (Table 34 and Table 3). Please rerun them at the reported width.

  4. The appendix settings disagree with the grid. The reported coefficient sits on the edge of the search box and one dataset reports a value outside it, the SNR table is computed at coefficients that match no reported row, and M.1 differs from the grid in the direction of one coefficient and in the channel count. Please say which range the reported settings came from, name the run behind each diagnostic, and align M.1.

  5. Which channels stay raw. With overlapping blocks there is no untouched middle, yet one appendix sentence says the middle channels are preserved. Please say which channels are raw in that case.

  6. The selection rule. The reported row is the best of the grid, and no section says the coefficients were chosen on validation. Please state the rule in Section 4.

  7. The stale value. Four tables still carry the superseded APAVA number. Please bring them to the corrected one.

  8. CCA placement. It wins on one dataset, and Appendix J still says its results are not in the main text. Please either cite it in the main text or drop the disclaimer.

  9. The reordering. Section 3.1 treats the physiological reordering as a stage that ran, Algorithm 1 marks it optional, and your reply gives array order as the default. Please say which it was.

  10. One mislabeled metric. Section 4.2 calls two numbers F1, and Table 5 gives them as precision. Please correct the sentence.

Where this leaves me

I keep my assessment at marginally below the acceptance threshold. Two things would move it up: the control in point 1 in the main text, and the corrections in points 7 and 10. If the control reproduces the gain, the mechanism claim has no source; if the method beats it at the reported settings, the mechanism survives, and I will say so either way.

AuthorsOpenReview ↗

Disentangling Structured Fusion from Invertible Reparameterization

New Q1: We thank the reviewer for this insightful comment. We agree that it is important to distinguish the benefits of CIF's structured channel interactions from those of a generic invertible linear reparameterization. To address this concern, we introduce an additional Fixed Random Orthogonal (FRO) baseline, which applies a randomly sampled orthogonal transformation along the channel dimension with CIF disabled.

Specifically, we independently sample a Gaussian random matrix and obtain an orthogonal matrix through QR decomposition. The resulting matrix is fixed throughout training and introduces no additional trainable parameters. To account for random initialization variability, we evaluate six independently sampled orthogonal matrices, each under five training seeds (41-45). We compare HMTCN, HMTCN+CIF, and HMTCN+FRO under identical backbone architectures, trainable parameter counts, optimizers, data splits, and experimental settings. Although FRO stores C2C^2 fixed matrix entries compared with CIF's two scalar coefficients, neither transformation introduces additional trainable parameters. Therefore, we match the trainable parameter counts to ensure a fair comparison in terms of optimization capacity.

The results show that CIF consistently outperforms both HMTCN and all six FRO variants across the four datasets. In particular, compared with the best-performing FRO configuration, CIF achieves accuracy improvements of 2.19, 5.85, 2.91, and 4.86 percentage points on APAVA, ADFTD, PTB, and TDBrain, respectively. These results demonstrate that a generic fixed random orthogonal transformation cannot reproduce the performance gains of CIF.

Overall, this experiment provides evidence that CIF's improvements are not merely attributable to applying an arbitrary fixed invertible transformation. Instead, they highlight the importance of its structured front-back channel interactions. We have included the experimental design, implementation details, and comprehensive results in the revised manuscript (Appendix C).

Table: Performance comparison of different fusion strategies under the subject-independent evaluation protocol.

DatasetModelAccuracyPrecisionRecallF1 scoreAUROCAUPRC
APAVAHMTCN82.49 ± 1.4082.38 ± 1.7981.20 ± 1.3281.60 ± 1.3991.10 ± 1.6391.30 ± 1.71
HMTCN + CIF86.54 ± 1.5586.33 ± 1.8886.10 ± 1.1186.08 ± 1.4694.43 ± 0.7094.57 ± 0.67
CIF (fro_seed_0)82.61 ± 1.8082.50 ± 2.3281.52 ± 1.6681.80 ± 1.7791.47 ± 1.2891.29 ± 1.40
CIF (fro_seed_1)80.63 ± 2.7881.03 ± 3.6679.14 ± 2.2379.54 ± 2.5789.89 ± 3.2889.92 ± 3.37
CIF (fro_seed_2)84.35 ± 2.2984.05 ± 2.0683.45 ± 2.8983.65 ± 2.6292.67 ± 2.1092.78 ± 2.07
CIF (fro_seed_3)81.87 ± 2.3882.84 ± 1.6679.61 ± 3.0680.39 ± 2.9990.80 ± 1.1890.95 ± 1.36
CIF (fro_seed_4)81.09 ± 4.6481.12 ± 5.3279.68 ± 4.3580.12 ± 4.6389.63 ± 4.0189.97 ± 3.63
CIF (fro_seed_5)83.33 ± 1.7883.19 ± 2.1082.14 ± 1.6582.52 ± 1.7692.04 ± 1.6092.15 ± 1.47
ADFTDHMTCN52.05 ± 2.2250.45 ± 3.0050.40 ± 2.5549.48 ± 2.7069.43 ± 2.8450.99 ± 3.15
HMTCN + CIF58.56 ± 0.9355.65 ± 0.8155.86 ± 0.7955.42 ± 0.8276.07 ± 0.5959.75 ± 0.67
CIF (fro_seed_0)52.17 ± 1.1249.09 ± 2.2649.18 ± 2.1148.20 ± 3.0268.62 ± 2.2550.11 ± 2.39
CIF (fro_seed_1)52.17 ± 2.2850.45 ± 3.4250.53 ± 4.4849.63 ± 4.6269.97 ± 2.9752.09 ± 3.51
CIF (fro_seed_2)52.71 ± 2.3650.57 ± 3.3350.44 ± 3.6549.69 ± 3.7869.86 ± 2.9051.64 ± 3.98
CIF (fro_seed_3)52.57 ± 2.1950.74 ± 2.5151.06 ± 2.6550.54 ± 2.5870.20 ± 2.5752.05 ± 2.66
CIF (fro_seed_4)50.91 ± 3.1750.03 ± 3.0749.78 ± 2.8249.16 ± 3.0668.88 ± 2.6550.38 ± 3.55
CIF (fro_seed_5)52.45 ± 1.4250.92 ± 1.7050.06 ± 1.4049.75 ± 1.5869.54 ± 1.4751.44 ± 2.02

New Q7& Q10: Thank you very much for pointing this out. We have corrected it accordingly.

Reviewer_oWj1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper advocates for a data-centric approach to multichannel medical time series (MedTS) classification (EEG and ECG). Rather than designing complex spatial-temporal architectures (such as Graph Neural Networks or cross-channel Transformers), the authors propose Channel-Imposed Fusion (CIF), a lightweight input transformation that linearly combines two subsets of channels (x_(new) = aX + bY) using two shared scalar coefficients a and b. To evaluate CIF without confounding architectural complexity, the authors pair it with HMTCN (Hidden-layer Mixed Bidirectional Temporal Convolutional Network), which sums the outputs of forward and time-reversed TCN branches. Experiments across five datasets (APAVA, TDBRAIN, ADFTD, PTB, PTB-XL) under subject-independent splits compare CIF against 8 spatial preprocessing baselines and over a dozen time series classifiers.

Critical strengths

Practical Data-Centric Perspective & Computational Efficiency: Focusing on input representation quality rather than quadratic O(C2)\mathcal{O}(C^2) channel-mixing layers is well-motivated for small, low-SNR clinical cohorts. CIF introduces only O(1)\mathcal{O}(1) parameters (two scalars a, b) and O(n)\mathcal{O}(n) spatial overhead, enabling fast training and inference.

Comprehensive Benchmarking Under Subject-Independent Splits: The evaluation spans 5 medical datasets across EEG and ECG modalities using clinically realistic subject-independent splits. Including both classical spatial filters (CAR, Bipolar, Laplacian, CSP, ICA) and modern neural architectures (Medformer, MedGNN, PatchTST, Crossformer) is commendable.

Cross-Backbone Plug-and-Play Utility: Transferability experiments (Appendix M) demonstrate that applying CIF prior to existing architectures (e.g., EEGNet, Conformer, Autoformer, Crossformer) consistently improves their baseline classification metrics.

Critical weaknesses

For a premier representation learning venue like ICLR, linearly combining the first n channels and last n channels with two shared scalars is too elementary to stand as a primary methodological contribution.

Furthermore, the companion backbone (HMTCN) simply runs a standard dilated TCN forward and backward and adds the hidden features—a standard bidirectional TCN design that offers minimal architectural novelty.

The paper motivates CIF using anatomical symmetry and physiological coupling—specifically highlighting left-right symmetrical prefrontal electrodes (Fp1-Fp2) canceling ocular blink artifacts in EEG and Einthoven's triangle in ECG. Actual Pairing Mechanics: In Algorithm 1 and Appendix C.1 (Page 26), CIF does not pair homologous anatomical regions. Instead, it takes the default alphabetical/arbitrary channel array order from the Medformer repository and pairs X[:, :, :n] with X[:, :, -n:].

Appendix M.1 (Tables 22–24, Pages 48–65) contains 18 pages of manual grid searches over direction t ∈ {-1, 1}, block size n, coefficients a, b ∈ [-1, 1], and whether a or b are frozen or learnable. These are crazy numbers of hyper-params tuning, making it practically not usable.

Questions for authors
  1. Hyperparameter Sensitivity: how sensitive is the hyper-params?
  2. HMTCN Baseline Degradation: Why does combining the Forward and Backward TCN branches without CIF degrade accuracy compared to a single unidirectional TCN on TDBRAIN, ADFTD, and PTB (Table 17)?
3 comments

We sincerely thank you for your professional and insightful comments. We hope our response adequately addresses your concerns.

Q1: To investigate the hyperparameter sensitivity of CIF, we conduct an additional corner-case analysis by evaluating four combinations of the interaction coefficients, i.e., (a,b)∈{−1,+1}(a,b)\in\{-1,+1\}, while fixing the channel partition size to n=⌊C/2⌋n=\lfloor C/2\rfloor and retaining the dataset-specific default direction tt.

As shown in Table below, the optimal coefficient combination varies across datasets, indicating that different time-series datasets may benefit from different channel interaction patterns. Importantly, by selecting an appropriate configuration from these four representative settings, CIF consistently outperforms the HMTCN baseline across all four datasets. Specifically, the best configurations yield Accuracy improvements of 1.81, 1.70, 3.15, and 3.85 percentage points on APAVA, ADFTD, PTB, and TDBrain, respectively.

These results suggest that CIF benefits from dataset-specific hyperparameter optimization rather than relying on a universally optimal coefficient combination. The performance variations across different settings further demonstrate the flexibility of CIF in adapting channel interactions to dataset-specific characteristics. Therefore, although appropriate hyperparameter selection is beneficial, CIF consistently achieves improvements over the no-CIF baseline across all evaluated datasets when properly configured. We have added this analysis to the revised manuscript to provide a more comprehensive assessment of its hyperparameter sensitivity.

Moreover, hyperparameter tuning is not unique to CIF, as existing medical time-series models, such as Medformer and MedGNN, also involve model-specific hyperparameters that require appropriate adjustment to achieve optimal performance.

Table. Sensitivity of CIF to corner settings (a,b)∈{−1,+1}2(a,b)\in\{-1,+1\}^2 with fixed n=⌊C/2⌋n=\lfloor C/2\rfloor under the subject-independent evaluation protocol. HMTCN denotes the no-CIF baseline.

APAVA

ModelsAccuracyPrecisionRecallF1 scoreAUROCAUPRC
HMTCN82.49±1.4082.38±1.7981.20±1.3281.60±1.3991.10±1.6391.30±1.71
HMTCN + CIF (+1,+1)(+1,+1)84.30±1.2183.80±1.2483.79±1.3883.76±1.2892.53±0.7992.67±0.77
HMTCN + CIF (+1,−1)(+1,-1)77.08±1.7877.30±2.8275.08±1.1275.56±1.3885.06±0.7385.25±1.06
HMTCN + CIF (−1,+1)(-1,+1)77.04±1.4677.68±2.2574.54±0.9775.16±1.1684.76±1.1985.27±1.21
HMTCN + CIF (−1,−1)(-1,-1)84.12±1.0584.18±1.5982.90±0.9583.32±1.0192.68±0.9392.78±0.89

ADFTD

ModelsAccuracyPrecisionRecallF1 scoreAUROCAUPRC
HMTCN52.05±2.2250.45±3.0050.40±2.5549.48±2.7069.43±2.8450.99±3.15
HMTCN + CIF (+1,+1)(+1,+1)53.75±0.7149.43±1.0249.35±0.4848.82±0.8269.36±0.7751.47±1.07
HMTCN + CIF (+1,−1)(+1,-1)52.47±3.0250.78±3.1650.59±2.8450.16±2.9170.05±2.6352.40±3.11
HMTCN + CIF (−1,+1)(-1,+1)53.25±4.4652.42±3.1251.22±3.1650.77±3.5271.77±3.2254.40±4.58
HMTCN + CIF (−1,−1)(-1,-1)53.35±1.5349.43±0.7549.65±0.5449.05±1.0369.05±0.6051.18±0.44

PTB

ModelsAccuracyPrecisionRecallF1 scoreAUROCAUPRC
HMTCN81.87±1.8786.50±1.2473.49±2.9075.84±3.2094.20±0.2993.04±0.45
HMTCN + CIF (+1,+1)(+1,+1)83.86±3.5088.57±1.6876.28±5.4878.61±5.9593.57±4.4593.18±4.22
HMTCN + CIF (+1,−1)(+1,-1)81.05±0.9985.47±0.8472.42±1.5774.69±1.7693.69±0.9592.05±0.90
HMTCN + CIF (−1,+1)(-1,+1)79.23±2.3284.91±2.2469.43±3.4271.19±4.1192.80±1.3791.34±1.51
HMTCN + CIF (−1,−1)(-1,-1)85.02±1.3588.50±1.0178.27±1.9980.88±1.9995.79±0.7495.11±0.72

TDBrain

ModelsAccuracyPrecisionRecallF1 scoreAUROCAUPRC
HMTCN84.90±2.6086.02±2.0084.90±2.6084.76±2.7493.94±1.9294.20±1.85
HMTCN + CIF (+1,+1)(+1,+1)88.75±2.7589.24±2.2788.75±2.7588.70±2.8096.45±1.3296.56±1.27
HMTCN + CIF (+1,−1)(+1,-1)85.44±2.8086.04±2.3085.44±2.8085.36±2.8994.07±1.7994.30±1.69
HMTCN + CIF (−1,+1)(-1,+1)85.40±1.5786.08±1.3385.40±1.5785.32±1.6394.10±1.3794.36±1.32
HMTCN + CIF (−1,−1)(-1,-1)87.73±2.7588.19±2.5387.73±2.7587.69±2.7895.07±2.2595.14±2.27

Q2:

We thank the reviewer for raising this insightful question. We believe that this phenomenon highlights an important characteristic of medical time-series classification: increasing model capacity does not necessarily improve performance, particularly under low signal-to-noise ratio (SNR) conditions.

On datasets such as TDBRAIN, ADFTD, and PTB, medical signals often exhibit substantial noise, inter-channel interference, and physiological variability. Although the Forward and Backward TCN branches capture temporal dependencies from complementary directions, simply adding their hidden representations does not guarantee effective information integration. Under low-SNR conditions, both branches may capture unreliable or redundant responses, and their direct summation can further propagate these undesirable components, potentially degrading the overall representation quality compared with a single unidirectional TCN.

In contrast, on relatively clean datasets such as FLAAP and UCI-HAR, the bidirectional branches can more effectively exploit complementary temporal contexts, making their combination more beneficial and resulting in improved performance over the unidirectional baseline.

This contrast reveals that the effectiveness of increased model capacity is closely related to the intrinsic quality of the input data. In particular, medical time-series classification is not merely a problem of extracting richer temporal features, but also of effectively handling noisy and heterogeneous signal representations. This observation further motivates our data-centric perspective: rather than simply increasing architectural complexity, improving the quality of data representations is essential for achieving reliable classification under challenging medical signal conditions.

Q3: Anatomical Motivation and Actual Pairing Mechanism

First, regarding the distinction between anatomical motivation and the actual pairing mechanism, the default implementation of CIF does not explicitly establish anatomically homologous channel pairs, such as Fp1–Fp2 in EEG. Specifically, following the original Medformer implementation, CIF partitions the input channels according to their existing array order and performs interactions between the first nn channels and the last nn channels.

It is important to emphasize that the primary objective of CIF is to explore structured cross-channel interaction, rather than to strictly constrain channel interactions to anatomically corresponding pairs. To further investigate the role of physiological structural priors in cross-channel interaction, Appendix J includes additional experiments based on specific physiological structures, including an experiment that uses only Fp1 and Fp2, two EEG channels exhibiting left-right hemispheric symmetry, for interaction. These supplementary experiments further examine the effectiveness of CIF from the perspective of specific physiological structures. Therefore, the examples of anatomical symmetry and physiological coupling in the original manuscript are intended to provide physiological insights into the design of cross-channel interaction, rather than to describe the explicit pairing strategy adopted by the default CIF implementation.

CCA-Based Simplification and Hyperparameter Complexity

Second, regarding hyperparameter complexity, to reduce the additional tuning overhead associated with the fusion coefficients aa and bb, we further investigate a simplified CCA-based formulation of CIF. Specifically, this formulation eliminates the need to manually tune the two fusion coefficients aa and bb, as well as to determine whether they should be frozen or learnable. Instead, it only requires specifying the channel grouping size nn, thereby substantially simplifying the configuration process and reducing the hyperparameter search overhead.

To verify the effectiveness of this simplified formulation, we incorporate CIF(CCA) into HMTCN and conduct additional experiments on five physiological time-series datasets. The results are presented in the following table.

DatasetsModelsAccuracyPrecisionRecallF1 scoreAUROCAUPRC
APAVAHMTCN82.49±1.4082.38±1.7981.20±1.3281.60±1.3991.10±1.6391.30±1.71
HMTCN + CIF(CCA)87.91±1.6089.81±0.9385.85±2.1286.93±1.9196.97±0.5296.91±0.49
ADFTDHMTCN52.05±2.2250.45±3.0050.40±2.5549.48±2.7069.43±2.8450.99±3.15
HMTCN + CIF(CCA)53.31±1.4049.90±1.5550.21±1.5749.75±1.6370.08±1.2351.96±1.92
PTBHMTCN81.87±1.8786.50±1.2473.49±2.9075.84±3.2094.20±0.2993.04±0.45
HMTCN + CIF(CCA)84.49±4.8488.07±2.7877.59±7.6079.63±7.2494.27±2.3293.45±2.44
TDBrainHMTCN84.90±2.6086.02±2.0084.90±2.6084.76±2.7493.94±1.9294.20±1.85
HMTCN + CIF(CCA)91.48±2.5391.67±2.5291.48±2.5391.47±2.5397.47±1.5397.56±1.47
PTB-XLHMTCN72.92±0.8864.21±1.5660.60±0.7261.49±0.8289.71±0.7266.44±1.49
HMTCN + CIF(CCA)73.08±0.4464.10±0.9960.57±0.9361.62±0.6089.92±0.4666.71±0.73

As shown in the table, CIF(CCA) consistently improves accuracy across all five datasets, while also achieving overall improvements in AUROC and AUPRC. Specifically, accuracy increases by 5.42 and 6.58 percentage points on APAVA and TDBrain, respectively, and by 2.62 and 0.16 percentage points on PTB and PTB-XL, respectively. On ADFTD, accuracy and AUROC also improve by 1.26 and 0.65 percentage points, respectively. Although the magnitude of improvement varies across datasets, these results demonstrate that the effectiveness of CIF's cross-channel interaction does not necessarily depend on extensive fusion-coefficient tuning. Its simplified formulation, which only requires specifying the channel grouping size, can still maintain effective performance across diverse physiological time-series datasets.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Newly Added Content & Revised Content

To All Reviewers

We use red to highlight newly added content and blue to indicate revised content in the main manuscript.

Newly Added Content:

  1. In Section 3.1, we clarify that only two interaction coefficients, aa and bb, are selected for hyperparameter tuning to facilitate efficient hyperparameter optimization while keeping the configuration space manageable.

  2. In Section 4, we explicitly clarify that the training, validation, and test sets strictly follow the same data-splitting protocol as Medformer, ensuring a consistent and fair evaluation setting.

  3. In Section 4.2 (2), we discuss that directly transferring CIF hyperparameters optimized for HMTCN to other backbones may lead to suboptimal performance or even negative transfer. Nevertheless, appropriate hyperparameter adjustment can still improve the performance of these backbones.

  4. In Appendix O.4, we conduct an additional hyperparameter sensitivity analysis. Specifically, we fix n=⌊C/2⌋n=\lfloor C/2\rfloor and evaluate four combinations of interaction coefficients, i.e., (a,b)∈{−1,+1}(a,b)\in\{-1,+1\}. The results demonstrate that at least one of the four configurations consistently improves upon the corresponding baseline across all four datasets.

Revised Content:

  1. In Section 3.1, we correct the description of the CIF fusion process to clarify that the original features are directly replaced rather than concatenated.

  2. In Table 3, we correct the best-performing result reported for the APAVA dataset.

AuthorsOpenReview ↗

New Insights for the Research Community

We sincerely thank the reviewer for raising this important concern regarding the methodological novelty of CIF and HMTCN. We would like to clarify that the primary objective of this work is not to introduce another sophisticated network architecture, but to investigate an underexplored yet fundamental question in medical time-series learning: Can improving the quality of data representations be more effective than continuously increasing model complexity?

1. Rethinking the Problem: From Model-Centric to Data-Centric Learning

Recent advances in medical time-series classification have largely focused on increasingly sophisticated model architectures, with performance improvements often pursued through more elaborate feature extraction and fusion mechanisms. However, an equally important challenge lies in the intrinsic nature of medical time-series data itself. In particular, medical signals often exhibit low signal-to-noise ratios, substantial inter-channel interference, and complex physiological dependencies. Under such conditions, increasing model capacity does not necessarily translate into better representations or more reliable classification.

Our motivation is therefore fundamentally different from that of conventional model-centric approaches. Rather than introducing additional architectural complexity to enhance the model's representation capacity, we investigate whether explicitly exploiting inter-channel relationships can improve the quality of the representations presented to the classifier. From this perspective, CIF is designed as a simple, plug-and-play, and model-agnostic data-centric strategy that enhances channel interactions without requiring modifications to the underlying backbone architecture. Its simplicity is intentional: we aim to demonstrate the effectiveness of addressing data representations themselves, rather than attributing performance improvements to increasingly complicated network designs.

2. On the Simplicity of CIF and HMTCN

We agree that CIF employs a lightweight linear fusion mechanism and that HMTCN follows a standard bidirectional TCN design. However, this simplicity is not a limitation of our approach, but rather an intentional design philosophy. Our goal is to develop a simple yet effective method that demonstrates the potential of improving medical time-series representations through better exploitation of intrinsic inter-channel relationships, rather than relying on increasingly sophisticated network architectures.

Specifically, CIF captures and exploits inter-channel relationships through a lightweight and shared fusion mechanism, providing an effective means of enhancing data representations without introducing substantial architectural complexity. Meanwhile, we deliberately adopt a relatively simple backbone, HMTCN, to provide a clear experimental setting in which the contribution of CIF can be independently investigated. This design also allows us to demonstrate that meaningful performance improvements can be achieved without relying on elaborate backbone architectures.

More importantly, the plug-and-play nature of CIF enables its integration into different backbone architectures, suggesting that its effectiveness is not inherently tied to a particular model design. We believe that the key insight this work offers to the community is that improving how existing models exploit the intrinsic structure of medical time-series data can be a simple yet effective alternative to continuously increasing model complexity. Rather than pursuing complexity for its own sake, we advocate a complementary data-centric perspective that places greater emphasis on the quality of data representations.

3. Beyond Architectural Complexity: Empirical Evidence and Broader Implications

More importantly, our findings provide empirical evidence for a broader perspective. By combining CIF with a simple TCN backbone, we achieve competitive, and in several cases superior, performance compared with more sophisticated Transformer-based approaches across multiple medical time-series datasets. These results suggest that stronger classification performance does not necessarily require more complex architectures. Instead, improving the quality of input representations through effective exploitation of the intrinsic data structure can be an equally powerful and potentially more transferable direction.

We do not argue that model-centric innovations are unnecessary, nor do we claim that simple models universally outperform complex ones. Rather, we hope our results encourage the community to reconsider the relative importance of data-centric and model-centric strategies in medical time-series learning. The central message of this work is that progress should not be measured solely by architectural complexity; understanding and effectively exploiting the characteristics of medical data can be equally important.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.