acceptodds
Under review as a conference paper at ICLR 2027

Meredith: Learning Physiological Rhythms and Delayed Channel Interactions for Medical Time Series Classification

Abstract

Accurate medical time series classification is important for identifying pathological conditions and supporting clinical diagnosis and disease management. Recent deep learning methods have achieved remarkable progress in this task, with their core designs largely centered on temporal representation learning and cross-channel modeling. However, prevailing strategies, such as generic multi-scale modeling for temporal representations and attention mechanisms or graph learning for cross-channel relationships, often lack explicit consideration of the intrinsic characteristics of physiological signals, potentially limiting the identification of diagnostically relevant patterns. To address this limitation, we draw upon physiological characteristics of medical time series to guide temporal representation learning and cross-channel modeling. First, we introduce Rhythm Mixer, which incorporates physiological rhythms into representation learning through learnable frequency decomposition and adaptive frequency-band aggregation, enabling medical-specific temporal representations beyond generic multi-scale modeling. Second, motivated by the temporal delays that can arise from the propagation and interaction of physiological activities, we propose the Delay-Aware Channel Enhancement Network (DACE Net), which accounts for different temporal delays to enhance each channel representation with delay-aware information from other channels. Building around these two core components, we develop Meredith, a physiology-informed framework for medical time series classification. Extensive experiments on five medical time series benchmarks demonstrate the strong and robust performance of Meredith against 10 competitive baselines under both Fixed Subject Split and Monte Carlo Cross-Validation protocols.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Medical Time Series Classification, Deep Learning

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_feweOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This manuscript presents Meredith, a physiology-informed framework for medical time series classification. The framework consists of two key components. One of them is Rhythm Mixer, which learns rhythm-aware representations through frequency decomposition and adaptive aggregation. The other is DACE Net, which incorporates temporal delays to enhance cross-channel representations. The authors evaluate the proposed method against 10 competitive baselines on medical time series benchmarks, demonstrating promising classification performance.

Critical strengths

1、 The paper is well motivated by the characteristics of physiological signals, particularly physiological rhythms and temporally delayed cross-channel relationships. The overall research motivation is clear and relevant to medical time series classification. 2、 The experimental evaluation is relatively comprehensive, covering multiple datasets and a wide range of baseline methods. The authors also adopt both Fixed Subject Split and Monte Carlo Cross-Validation protocols. The main experimental results demonstrate competitive performance across several benchmarks.

Critical weaknesses

1、 The core delay-aware aggregation in DACE Net, under its current parameterization, appears mathematically equivalent to a constrained multichannel causal convolution. The paper does not sufficiently clarify the methodological novelty of this formulation or establish its advantages over conventional temporal convolutions through controlled comparisons. 2、 Some ablation results exhibit unexpected behavior that is not adequately discussed. For example, on the PTB dataset, removing both Rhythm Mixer and DACE Net (w/o Both) achieves higher Accuracy and F1 scores than removing either component individually. This non-monotonic behavior raises questions about the interactions between the proposed modules. 3、 Appendix A.1 indicates that some baseline results are taken from previously published studies, including TeCh, rather than obtained by reimplementing all methods under a unified experimental setting. Although using published results is acceptable, the paper should more clearly establish whether the preprocessing procedures, subject splits, label definitions, and evaluation protocols are consistent across the compared methods. 4、 Table 6 reports 72 subjects for the TDBrain dataset, whereas Appendix A.2.1 describes the selection of 25 patients with Parkinson's disease and 25 healthy controls, totaling 50 subjects. This discrepancy raises concerns about the clarity of the dataset description and experimental setup.

Questions for authors

1、 How well does DACE Net generalize to subject-specific or time-varying physiological delays? 2、 Does Rhythm Mixer benefit from physiological frequency priors rather than generic learnable frequency decomposition? Would randomly initialized or uniformly spaced frequency bands achieve comparable performance? 3、 How does the computational and parameter complexity of DACE Net scale with the number of channels and the maximum temporal delay?

2 comments
AuthorsOpenReview ↗

Response to Reviewer fewe (Part 1)

Dear Reviewer fewe:

Thank you for your valuable feedback. We have carefully addressed your comments, and our detailed responses are provided below.

W1. Mathematical Interpretation and Advantages of DACE Net over Conventional Convolutions

A1:

We thank the reviewer for this insightful comment. We acknowledge that the delay-aware aggregation in DACE Net can be mathematically interpreted as a constrained multichannel causal convolution. However, the key distinction lies in its structured parameterization, which explicitly separates cross-channel connectivity from temporal delay distributions.

(i). Methodological Distinction.

Unlike conventional convolutions that directly learn unconstrained spatiotemporal kernels, DACE Net parameterizes the effective convolution weights as Wi,j,d=Ai,jPi,j,dW_{i, j, d}=A_{i, j} P_{i, j, d}, where Ai,jA_{i,j} represents the learned connectivity between channels ii and jj, and Pi,j,dP_{i,j,d} characterizes their temporal delay distribution at delay d∈{0,…,τ}d \in \{0,\ldots,\tau\}, with τ\tau denoting the maximum temporal delay. Both are normalized through Softmax, yielding non-negative, normalized aggregation weights. This design introduces an explicit structural inductive bias for modeling delayed cross-channel interactions, rather than proposing a fundamentally new convolution operation.

(ii). Controlled Comparisons with Conventional Convolutions.

To evaluate the effectiveness of this design, we conducted additional experiments replacing the core delay-aware aggregation with either Causal Conv1D or Standard Conv1D, while keeping the remaining architecture and training settings unchanged. As shown in the following table, DACE Net consistently outperforms both alternatives, demonstrating the effectiveness of its structured delay-aware parameterization.

DatasetModel VariantsAccuracyPrecisionRecallF1 scoreAUROCAUPRC
TDBrainDACE Net94.9895.1094.9894.9799.1099.11
Causal Conv1D88.5689.7289.5688.5097.1597.17
Standard Conv1D88.9789.4588.9788.9596.1896.23
APAVADACE Net91.7093.1290.2391.1298.0298.08
Causal Conv1D82.6780.8479.5881.0396.2596.30
Standard Conv1D82.8181.1279.7081.1796.4496.47

W2. Non-monotonic Ablation Results and Module Interactions

A2:

We acknowledge the non-monotonic behavior on PTB and provide the following clarification.

(i). Explanation of Non-monotonic Behavior.

Rhythm Mixer and DACE Net capture complementary information through physiological rhythm modeling and delay-aware channel enhancement, respectively. Their outputs are integrated before the Transformer encoder, meaning that their contributions are not necessarily additive. Moreover, each ablation variant is independently trained from scratch, and removing different modules changes both the feature representations and the optimization process. Therefore, removing both modules may occasionally yield better performance than removing either one individually.

(ii). Effectiveness of Joint Modeling.

As shown in Table 2, although the "w/o Both" variant outperforms the two individual ablations on PTB in terms of Accuracy and F1 score, the complete Meredith model substantially outperforms all three variants across all six evaluation metrics. Furthermore, "w/o Both" variant performs worse than both individual ablations on the other four datasets.

These findings indicate that the contributions of Rhythm Mixer and DACE Net are not strictly additive, while the consistent superiority of the complete model across all five datasets supports the effectiveness of their joint integration.

W3. Fairness and Consistency of Baseline Comparisons

A3:

(i). Fixed Subject Split.

Our experiments follow the evaluation protocol established by Medformer[1], which is also adopted by MedGNN[2] and TeCh[3]. Specifically, the preprocessing procedures, subject splits, and label definitions are consistent across these studies, ensuring fair comparisons under the same experimental setting. Therefore, directly adopting previously reported baseline results is appropriate and avoids unnecessary reimplementation. Notably, TeCh follows the same practice, explicitly stating on page 6 that "The results of all baselines on the five MedTS datasets are directly taken from Medformer."

(ii). Monte Carlo Cross-Validation.

Please refer to A3(i) to Reviewer KDrH for clarification of the evaluation protocol and baseline comparisons under this setting.

We will further clarify these details and the sources of the baseline results in Appendix A.1 to improve transparency and reproducibility.

AuthorsOpenReview ↗

Response to Reviewer fewe (Part 2)

W4. Clarification of the Number of Subjects in TDBrain Dataset

A4:

We thank the reviewer for pointing out this inconsistency. We apologize for the reporting error in Table 6, where the number of subjects in TDBrain was incorrectly reported as 72 instead of 50. We will correct this in the revised manuscript.

Importantly, this error does not affect the experimental setup or the fairness of our comparisons. Our experiments follow the same TDBrain preprocessing protocol as Medformer, which explicitly states in Appendix B.2: "In this paper, we utilize a subset of this dataset containing 25 subjects with Parkinson’s disease and 25 healthy controls, all under the eye-closed task condition." Therefore, our experiments consistently use the same 50-subject subset as Medformer, and the discrepancy is limited to the reporting error in Table 6.

Q1. Generalization of DACE Net to Subject-Specific and Time-Varying Delays

A5:

(i). Generalization across Subjects.

DACE Net learns channel-pair-specific delay distributions shared across subjects, rather than explicitly estimating individualized delays. Nevertheless, Meredith is evaluated under both Fixed Subject Split and Monte Carlo Cross-Validation protocols, where training and test subjects are disjoint. The results demonstrate that the learned representations generalize to unseen subjects, although they do not directly establish subject-specific delay estimation.

(ii). Adaptation to Time-Varying Delays.

The current DACE Net uses globally learned delay distributions that remain fixed during inference. While its downstream refinement layers can produce input-dependent representations, the delay weights themselves do not adapt dynamically to temporal changes. We acknowledge this limitation and consider input-dependent delay estimation a promising direction for future work.

Q2. Contribution of Physiological Frequency Priors.

A6:

Please refer to A1(i) to Reviewer KDrH, where we provide controlled comparisons among Canonical (ours), Uniform, Random, and Fixed Canonical Rhythm Bank variants. The results consistently demonstrate the benefits of physiologically informed frequency initialization and adaptive frequency learning.

Q3. Computational and Parameter Complexity of DACE Net

A7:

The complexity of DACE Net depends on the number of channels CC, sequence length TT, maximum temporal delay τ\tau, and number of refinement layers LrL_{r}.

(i). Parameter Complexity.

DACE Net employs C2C^{2} channel weights and C2(τ+1)C^{2}(\tau+1) delay weights, together with a projection layer and LrL_{r} channel refinement layers. The total number of learnable parameters is Nparams =C2(τ+Lr+3)+(3Lr+5)CN_{\text {params }}=C^{2}\left(\tau+L_{r}+3\right)+\left(3 L_{r}+5\right) C (including the learnable parameters of BatchNorm layers), yielding a parameter complexity of O(C2τ)O\left(C^{2} \tau\right) when LrL_{r} is fixed.

(ii). Computational Complexity.

Delay-aware aggregation requires O(TC2(τ+1))O\left(T C^{2}(\tau+1)\right) operations, while channel aggregation and subsequent projection and refinement require O(TC2(Lr+2))O\left(T C^{2}\left(L_{r}+2\right)\right) operations. Therefore, the overall computational complexity is O(TC2(τ+Lr+3))O\left(T C^{2}\left(\tau+L_{r}+3\right)\right), which simplifies to O(TC2τ)O\left(T C^{2} \tau\right) for a fixed LrL_{r}.

Thus, both parameter and computational complexity scale quadratically with the number of channels and linearly with the maximum temporal delay. This is consistent with the asymptotic complexity of conventional multichannel causal convolution, while DACE Net introduces structured channel connectivity and delay distributions.

References:

[1] Medformer: A Multi-Granularity Patching Transformer for Medical Time-Series Classification (NeurIPS 2024)

[2] Towards multi-resolution spatiotemporal graph learning for medical time series classification (WWW 2025)

[3] Decentralized Attention Fails Centralized Signals: Rethinking Transformers for Medical Time Series (ICLR 2026)

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

This paper introduces Meredith, a framework for medical time series classification that emphasizes the incorporation of physiological priors into representation learning. The two main components of Meredith are the Rhythm Mixer and the Delay-Aware Channel Enhancement Network (DACE Net). The Rhythm Mixer decomposes the input with learnable Gaussian frequency masks initialized at known EEG bands or ECG waveform frequency ranges, processes each band separately, and combines using softmax. DACE Net learns per-channel-pair weights over temporal delays, motivated by conduction delays in the heart, and delayed interactions across brain regions. Meredith achieves strong performance across several EEG and ECG benchmarks, and ablations demonstrate the contribution of major design choices.

Critical strengths
  1. The motivation is clear and appealing: physiological signals have known structure (canonical EEG rhythms, ECG waveform frequencies, propagation delays across channels) that generic time-series architectures do not capitalize on, and building this structure into the model as inductive bias seems reasonable.

  2. The methodology is simple and well explained. The design of Rhythm Mixer and DACE Net are easy to follow from the text and Figure 1, and should be straightforward to reproduce.

  3. Empirical results are strong. Meredith is best or second-best across five EEG and ECG benchmarks under two subject-independent protocols, and generalizes to HAR datasets.

  4. The method is parameter-efficient, using far fewer parameters than other medical-specific baselines while matching or beating them.

Critical weaknesses
  1. The paper does not demonstrate that the physiological content of the priors is the primary cause of the strong performance. The ablations only remove entire components, which removes frequency decomposition and cross-channel modeling entirely, despite the stated contribution being the use of “physiological characteristics of medical time series to guide temporal representation learning and cross-channel modeling.”

  2. The HAR results are strong despite IMU data having no canonical band structure (the Rhythm Bank used is not stated), and the learned channel weights in Figure 10 are nearly uniform (0.061 to 0.065 on APAVA). The learned band centers and widths after training are never reported.

  3. If the physiological priors do not contribute beyond a good initialization, the contribution reduces to a learnable filter bank plus a lag-aware channel mixer, which has close prior work (e.g. SincNet-style learnable band-pass filters, lag modeling in Autoformer) and limits the novelty.

  4. Nearly all baseline numbers are copied from the TeCh paper rather than re-run, the source of the Table 7 baselines is not stated, and several baselines shift by 10 to 20 points between protocols on the same data (iTransformer on APAVA: 74.6 vs 95.7). This makes the fixed-split margins harder to interpret.

Questions for authors
  1. Can the authors isolate the contribution of the physiological content of the priors specifically? Controls such as initializing the Rhythm Bank at evenly spaced bands, freezing the band parameters at the canonical values, or running DACE Net with τ=0 would help, and reporting the learned band centers and widths after training would show whether the prior is retained. Any other analysis that separates the physiological initialization from the mechanism itself is fine too.

  2. What Rhythm Bank was used for the ECG and HAR datasets? Neither is specified. If the HAR bank was arbitrary, how do the authors reconcile the strong HAR results with the claim that physiological initialization matters?

  3. Where do the Table 7 baseline numbers come from? Appendix A.1 only covers Tables 1 and 8. Relatedly, was hyperparameter selection done on the validation split or the test split?

6 comments
AuthorsOpenReview ↗

Response to Reviewer KDrH (Part 1)

Dear Reviewer KDrH:

We sincerely thank you for your thoughtful comments. We have carefully addressed the concerns you raised. Our responses to each point are presented below.

W1 & W3 & Q1. Contribution of Physiological Priors and Methodological Novelty

A1:

(i). Contribution of Physiological Frequency Priors.

To isolate the contribution of physiological frequency priors, we conduct controlled experiments with three Rhythm Bank variants: Uniform, Random, and Fixed Canonical. Uniform and Random replace the physiologically initialized center frequencies with evenly spaced and randomly sampled values, respectively, while retaining learnable frequency parameters. Fixed Canonical preserves the predefined physiological bands without updating their parameters. All variants retain the same Rhythm Mixer architecture.

As shown in the following table, Canonical (ours) consistently outperforms all three variants across the three evaluated datasets. The superior performance over Uniform and Random demonstrates the effectiveness of physiologically informed frequency initialization, while the improvements over Fixed Canonical highlight the additional benefit of adaptive frequency learning. These results confirm that physiological priors contribute to classification performance beyond the learnable frequency decomposition mechanism itself.

DatasetRhythm VariantTrainableAccuracyPrecisionRecallF1 scoreAUROCAUPRC
AVAPACanonical (ours)Yes91.7093.1290.2391.1298.0298.08
UniformYes83.1684.5281.1382.1288.1288.63
RandomYes76.8778.5673.9674.8876.8476.92
Fixed CanonicalNo87.7589.3286.9389.5092.7492.78
ADFTDCanonical (ours)Yes54.8854.1954.6554.0572.5555.17
UniformYes46.5544.4944.2144.1664.1844.04
RandomYes44.0441.9940.4839.8862.6840.73
Fixed CanonicalNo53.5551.2151.0551.0770.3751.16
PTBCanonical (ours)Yes87.5090.6481.7684.3996.7296.40
UniformYes82.7185.0677.4479.8589.7089.40
RandomYes79.5084.8569.9271.9191.9590.50
Fixed CanonicalNo83.2285.7377.7080.1691.0290.71

(ii). Learned Rhythm Bank Analysis.

We further examine the learned center frequencies and bandwidths of the Rhythm Bank on APAVA and TDBrain. As shown in the following table, the learned parameters remain close to their physiological initialization, with mean absolute changes in center frequency and bandwidth of only 0.0177 and 0.0097 Hz on APAVA, and 0.0719 and 0.0635 Hz on TDBrain, respectively. These results indicate that Rhythm Mixer largely preserves the canonical rhythm structure while allowing modest dataset-specific adaptations.

DatasetRhythmInitial CenterLearned CenterInitial WidthLearned Width
APAVADelta2.25002.26611.75001.7591
Theta6.00005.97772.00001.9969
Alpha10.500010.51892.50002.5017
Beta21.500021.50768.50008.4787
Gamma37.500037.52347.50007.5134
TDBrainDelta2.25002.31861.75001.7358
Theta6.00006.06342.00001.9375
Alpha10.500010.53172.50002.5463
Beta21.500021.62388.50008.3632
Gamma37.500037.57197.50007.4425

(iii). Contribution of Temporal Delays in DACE Net.

To isolate the contribution of temporal delay modeling, we further evaluate DACE Net with τ=0\tau=0, where τ\tau denotes the maximum temporal delay. This variant retains cross-channel aggregation but excludes nonzero temporal delays. As shown in the following table, the original setting (τ=8\tau=8) consistently outperforms τ=0\tau=0 across all classification metrics on ADFTD and PTB. These results demonstrate that explicitly incorporating temporally delayed cross-channel information provides additional benefits beyond synchronous channel aggregation.

Datasetτ\tauAccuracyPrecisionRecallF1 scoreAUROCAUPRC
ADFTD854.8854.1954.6554.0572.5555.17
047.9548.8548.7449.7465.3247.49
PTB887.5090.6481.7684.3996.7296.40
083.4286.8376.4879.0892.5392.20
AuthorsOpenReview ↗

Response to Reviewer KDrH (Part 2)

A1:

(iv). Distinction from Existing Methods.

SincNet[1] and Autoformer[2] were originally proposed for speaker recognition and time series forecasting, respectively, whereas Meredith focuses on medical time series classification. Beyond this difference in application domains, these methods also differ in their modeling objectives and mechanisms.

(1) Rhythm Mixer vs. SincNet

SincNet learns parameterized band-pass filters for speech representation learning. In contrast, Rhythm Mixer incorporates signal-specific physiological rhythms into frequency decomposition, followed by independent rhythm-specific encoding and sample-adaptive aggregation to capture discriminative physiological patterns.

(2) DACE Net vs. Autoformer

Autoformer employs an Auto-Correlation mechanism to capture periodic temporal dependencies for forecasting. In contrast, DACE Net explicitly models temporally delayed cross-channel interactions through channel-pair-specific delay weights and subsequent channel aggregation and refinement, targeting delay-dependent relationships in physiological signals.

Furthermore, our additional experiments demonstrate that physiologically informed frequency initialization and adaptive rhythm learning both contribute to classification performance, while explicitly modeling nonzero temporal delays provides additional benefits beyond synchronous channel aggregation. These findings further support the distinct contributions of Rhythm Mixer and DACE Net beyond generic frequency filtering and temporal delay aggregation.

W2. Generalization Beyond Medical Time Series and Channel Weight Interpretation.

A2:

(i). Generalization Beyond Medical Time Series.

We would like to clarify two points regarding the HAR experiments.

First, evaluating medical time series classification models on non-medical datasets follows the experimental practice introduced by Medformer[3] (Table 7) and subsequently adopted by TeCh[4] (Table 3). Following these studies, we evaluate Meredith on FLAAP and UCI-HAR to assess its generalization beyond medical time series.

Second, strong performance on non-medical datasets does not contradict the physiological motivation of Meredith. For both HAR datasets, Rhythm Mixer uses the EEG Rhythm Bank for initialization, without introducing HAR-specific frequency priors. Although IMU signals do not exhibit canonical EEG rhythms, they can still contain informative frequency patterns and asynchronous cross-channel relationships. The learnable frequency decomposition and delay-aware channel enhancement in Meredith can therefore capture useful patterns beyond physiological signals. These results demonstrate the broader applicability of the proposed modeling mechanisms, rather than suggesting that physiological priors are unnecessary for medical time series classification.

Importantly, the contribution of physiological initialization is independently supported by the controlled experiments presented in A1(i), where Canonical consistently outperforms Uniform and Random initialization across both EEG and ECG datasets.

(ii). Interpretation of Learned Channel Weights in DACE Net.

The channel weights in Figure 10 are indeed close to uniform. Since these weights are softmax-normalized across source channels (line 273), the uniform reference value is 1/C1/C (e.g., 0.06250.0625 for APAVA with C=16C=16).

However, channel weights represent only one component of DACE Net. As described in Section 4.3, cross-channel information is aggregated through both channel weights and channel-pair-specific delay weights. Thus, even when channel weights are nearly uniform, different channel pairs can still exhibit distinct temporal delay distributions.

Figures 8 and 9 visualize the learned delay weights at different temporal offsets. Furthermore, the additional experiments in A1(iii) show that incorporating nonzero temporal delays consistently improves classification performance over synchronous channel aggregation (τ=0\tau=0). These results support the contribution of delay-aware modeling, which cannot be fully characterized by the channel weights in Figure 10 alone.

(iii). Learned Rhythm Bank Analysis.

Please refer to A1(ii).

AuthorsOpenReview ↗

Response to Reviewer KDrH (Part 3)

W4. Baseline Sources and Evaluation Protocols.

A3:

(i). Reliability and Sources of Baseline Results.

(1) Fixed Subject Split.

We follow the same dataset preprocessing and subject splits as Medformer[3], MedGNN[5], and TeCh[4]. Notably, the baseline results in Table 2 of TeCh are directly adopted from Table 4 of Medformer. Therefore, we use these published results for comparison under the same evaluation protocol, ensuring consistency with prior studies.

(2) Monte Carlo Cross-Validation.

Since Medformer and TeCh do not report results under this protocol, we reproduce the baselines using their publicly available implementations. Specifically, we replace the fixed subject split with Monte Carlo Cross-Validation while keeping the original model architectures and training configurations unchanged. We report the mean and standard deviation over five independent subject splits in Table 7. This protocol complements Fixed Subject Split by evaluating generalization robustness across different subject partitions.

We will further clarify the sources of the baseline results under the Monte Carlo Cross-Validation setting in Appendix A.1.

(ii). Performance Variations Across Evaluation Protocols.

We had already observed the unexpectedly strong performance of iTransformer on APAVA dataset under Monte Carlo Cross-Validation before submission. To ensure the reliability of the reported results, we repeated the experiments and obtained consistent findings. While the underlying reasons for this performance difference remain unclear, a detailed investigation of iTransformer's behavior is beyond the scope of this work.

Q2. Rhythm Bank Configurations for ECG and HAR Datasets.

A4:

(i). Rhythm Bank Configurations.

For ECG datasets (PTB and PTB-XL), we initialize the Rhythm Bank with seven frequency bands: Very Low (0--1 Hz), Low (1--10 Hz), Mid (10--25 Hz), High (25--45 Hz), T-wave (0--10 Hz), P-wave (5--30 Hz), and QRS (8--50 Hz). These bands combine general spectral partitions with frequency ranges inspired by characteristic ECG waveform components. For both HAR datasets (FLAAP and UCI-HAR), we use the same five-band EEG Rhythm Bank (Delta, Theta, Alpha, Beta, and Gamma), without introducing HAR-specific frequency priors. We will clarify these configurations in the revised manuscript.

(ii). Generalization Beyond Medical Time Series.

As discussed in A2(i), the HAR experiments are intended to evaluate Meredith's generalization beyond medical time series, rather than to validate physiological priors. Although IMU signals do not exhibit canonical EEG rhythms, they can still contain informative frequency patterns and asynchronous cross-channel relationships that Meredith can adaptively model.

Importantly, the contribution of physiological initialization is independently supported by the controlled experiments in A1(i), where Canonical consistently outperforms Uniform and Random initialization across EEG and ECG datasets.

Q3. Baseline Sources and Hyperparameter Selection.

A5:

(i). Sources of Table 7 Baseline Results.

Please refer to A3(i).

(ii). Hyperparameter Selection.

Hyperparameters were selected based on the validation F1 score rather than test performance. Early stopping and best checkpoint selection were also determined by validation F1, without using test results for model selection.

References:

[1] Speaker Recognition from Raw Waveform with SincNet. 2018 IEEE spoken language technology workshop (SLT)

[2] Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting (NeurIPS 2021)

[3] Medformer: A Multi-Granularity Patching Transformer for Medical Time-Series Classification (NeurIPS 2024)

[4] Decentralized Attention Fails Centralized Signals: Rethinking Transformers for Medical Time Series (ICLR 2026)

[5] Towards multi-resolution spatiotemporal graph learning for medical time series classification (WWW 2025)

Reviewer_cra7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a front-end learnable module specifically designed for medical time series. The modules is mainly composed of Rhythm Mixer, which operates learnable frequency-band filtering, and DACE net, which enhance representation via learned aggregation of temporal delays. The authors evaluate their model against ten baselines on five EEG and ECG datasets using subject-independent evalutions and show competitive/superior performance.

Critical strengths

The authors identify the significance of frequency information and the delayed cross-channel information in EEG and ECG data, providing solid motivation for the proposed design of Rhythm Mixer and DACE net.

Critical weaknesses

In my opinion, the ablation results (Table 2) raise concern for the soundness of the results. According to table 2, removing one of Rythm Mixer and DACE net causes performance to drop below most of the baselines, and removing both leads to significantly lower results, and in the case of ADFTD and PTB-XL, near chance-level performance. It seems to be difficult to reconciliate the facts that: 1. Rythm Mixer, DACE and the backbone are structurally disentangled; 2. The backbone alone is very weak; 3. The full model on average outperforms all baselines.

Questions for authors

See Weakness.

1 comment
AuthorsOpenReview ↗

Response to Reviewer cra7

Dear Reviewer cra7:

We sincerely appreciate your constructive feedback. We have carefully considered your concerns and provide our point-by-point responses below.

(i). Ablation Settings.

All ablation variants in Table 2 were independently trained from scratch, rather than obtained by removing components from a pretrained Meredith model. In particular, the w/o Both variant retains the Temporal Embedding, Transformer Encoder with learnable-query cross-channel attention, and Classification Head. Thus, the reported results reflect the performance of independently trained architectures rather than disruptions to pretrained representations.

(ii). Complementary Roles of the Components.

Structural separability does not imply functional independence. Rhythm Mixer captures rhythm-specific temporal patterns, while DACE Net models delayed cross-channel interactions. The subsequent Transformer encoder further integrates these enhanced representations for classification. In contrast, the backbone alone directly projects raw signals into channel-wise representations without explicitly modeling these physiological characteristics. Therefore, the substantial performance gains of Meredith are consistent with the complementary roles of its components and do not contradict their structural separability.

(iii). Understanding the Backbone Performance.

Importantly, the backbone does not consistently perform near chance level. As shown in Table 2, the w/o Both variant achieves 81.62% Accuracy and 93.66% AUROC on PTB, as well as 81.77% and 80.32% AUROC on APAVA and TDBrain, respectively. Notably, these three datasets involve binary classification, whereas ADFTD and PTB-XL involve three and five classes, respectively (see Table 6).

To further investigate the severe degradation on PTB-XL, we examined the confusion matrices of the complete Meredith and the w/o Both variant. The results reveal that the w/o Both variant collapses to predicting the majority class for all test samples, resulting in 45.46% Accuracy but only 20.00% macro Recall. In contrast, the complete Meredith successfully identifies samples from all five classes, achieving 73.24% Accuracy and 62.55% macro Recall on the same test set. These findings suggest that multi-class discrimination can be particularly challenging for the backbone without explicit physiological feature modeling, while incorporating Rhythm Mixer and DACE Net helps alleviate this limitation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Follow-up on Our Rebuttal Responses

Dear Reviewers,

We sincerely thank you all for your time and effort in reviewing our paper and providing constructive feedback.

As the discussion phase will end in two days, we would be very grateful if you could kindly take a moment to review our rebuttal and let us know whether our responses have addressed your concerns.

We greatly appreciate your consideration and any further feedback you may have.

Best regards,
Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.