acceptodds
Under review as a conference paper at ICLR 2027

MOSAIC: A Differentiable Phase-Amplitude Coupling Transformer for Medical Time Series Classification

Abstract

Neural oscillations interact through phase-amplitude coupling (PAC): the amplitude of a fast rhythm is modulated by the phase of a slow one. PAC is an established clinical biomarker, yet transformers for medical time series largely ignore it, treating signals as undifferentiated tokens. Classical PAC measures are typically computed offline on hand-chosen bands, so the bands that determine which coupling is detectable lie outside the classification objective. We present MOSAIC (Multi-Oscillatory Signal Attention with Integrated Coupling), which makes PAC estimation end-to-end differentiable. A spectrally applied learnable Gaussian filterbank lets gradients reach the band parameters, an FFT-based Hilbert transform gives instantaneous amplitude and phase, and a soft phase-binned amplitude histogram keeps the shape of the modulation rather than a single scalar. The resulting PAC tokens feed a cross-band transformer that learns which coupling pairs matter. On three subject-independent EEG and ECG benchmarks, jointly learning the bands and the coupling representation raises macro-F1 over fixed-band PAC features by 1.2× to 1.6×, and MOSAIC attains the highest reported accuracy, recall, and macro-F1 on all three against eleven published transformer baselines.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Phase-amplitude coupling, Medical time series, EEG, ECG, Transformers

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_p5iyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors introduce MOSAIC, a differentiable medical time-series classifier that uses a relaxation of phase-amplitude coupling (PAC) features as the input. This is implemented through a Gaussian filterbank with learnable parameters fbf_b and σb\sigma_b, which contrasts the fixed bands usually used in PAC analysis. The PAC features are joined with a parallel waveform branch to use as input tokens for a transformer used for classification. The loss function includes a classification loss, an orthogonality penalty to ensure PAC tokens encode distinct information, and a smoothness penalty that works under the assumption that amplitude changes smoothly as phase moves. Results show strong performance, outperforming transformer baselines and fixed-feature PAC classifiers in accuracy, precision, recall, and F1 score on two EEG datasets and one ECG dataset.

Critical strengths
  • Grounded and Well-Motivated Method: The paper merges a well-known EEG biomarker in PAC with a differentiable deep learning and data-driven framework, introducing adaptable frequency bands to classify and interpret time-series medical data. The authors also include proper ablations against different PAC estimation methods.

  • Strong Results Outperforming Baselines: The authors present results that significantly outperformed transformer models and fixed PAC classifiers in three time-series datasets.

Critical weaknesses
  • Architecture Justification: The main novelty presented in this paper is the differentiable relaxation of PAC features. However, the deep model architecture decisions are unfounded and unclear. The authors use a transformer on PAC features for classification. Given that only 12 total PAC tokens are passed through the model, the architectural complexity is surprising. Ablation against a simple MLP is necessary to justify it.

  • Interpretability Claims: While I do understand that aggregate reliance helps with interpretability, no empirical evidence of this is present to back up the authors’ claims. I would like to see the final fbf_b and aggregate reliance values on at least one dataset. Additionally, adding the waveform stream as input is a large structural change that further blurs the interpretability claims. The classifier can heavily rely on the raw waveform stream, which reduces the necessity of the PAC tokens. This is evident on the results in table 4, with performance becoming comparable to the baselines without the waveform stream.

  • Generalization to Medical Time-Series Data: PAC features are popular with EEG data, but the authors claim that MOSAIC can be used on all medical time-series data. These claims should be narrowed down, or substantive evidence of PAC features being used on different medical time-series data such as ECG is necessary. Given that raw waveform streams are part of the input data and there is limited literature on PAC in ECG, a single dataset of empirical evidence is not sufficient.

Questions for authors
  1. Can you provide any results showing evidence of the interpretability claims made in section 3, given that it serves as the primary motivation for the orthogonality penalty and the use of self-attention?

  2. Can you provide ablations on the orthogonality and smoothness terms in the loss function? While properly justified in text, no insight regarding their impact on performance is shown.

  3. The adaptable fbf_b possibly removes clinical utility of the learned PAC features if they significantly diverge from any understood frequency bands. Is there any way to constrain fbf_b based on prior information, or is this purely done based on the initialization?

  4. I have strong concerns regarding the addition of the waveform stream to the input tokens, as it diminishes the utility of the PAC tokens. Can you provide evidence and justification on the advantages of using joint PAC and waveform stream tokens as opposed to just the latter, either for interpretability or performance?

Reviewer_uLynOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes MOSAIC, which makes phase-amplitude coupling (PAC) estimation differentiable. Learnable band filters and a soft phase-binned amplitude histogram turn each channel into 12 PAC tokens, and these are fused with a waveform branch.

Critical strengths
  1. Bringing an established clinical biomarker (PAC) into an end-to-end model is a reasonable motivation, and the paper correctly identifies non-differentiable hard phase binning as one technical obstacle.
  2. The authors evaluate MOSAIC on both EEG and ECG data, but the theoretical rationale for extending the proposed EEG-oriented band-coupling approach to ECG signals remains unclear.
Critical weaknesses
  1. The paper uses classical PAC estimators, SincNet-style learnable band-pass filters, a textbook FFT Hilbert transform, and kernel-based soft histogramming. The contribution is mainly their composition into one differentiable pipeline. In addition, the method is closer to learnable feature engineering than to a new representation-learning approach.
  2. The mechanistic claim is contradicted by the ECG results. PAC describes interactions between neural oscillations, and the delta to high-gamma bands have no physiological meaning for ECG. Yet the frequency-based PAC stream matters most on the ECG dataset, PTB., where removing PAC stream costs 11.09 macro-F1, versus 4.74 on APAVA and 5.92 on TDBrain. This undermines the paper’s central argument that the representation “encodes a physiological mechanism,” and it raises the same question for the EEG results.
  3. No visualization analysis of what is learned. For example, Eq. 16 defines an aggregate attention score ᾱ_pa explicitly “for interpretation,” but it is never reported. There are no learned fbf_b and σbσ_b values, no training trajectories and no class-wise PBA profiles. Without these, claims that the model “learns which coupling pairs matter” or departs from clinical bands “only where the data warrant” are unsupported.
  4. Nearly all baselines are general time series models, most of them designed for forecasting. The comparison includes no established EEG-specific models and no ECG-specific models. Without these, the claimed state-of-the-art performance on EEG and ECG benchmarks is not convincing.
  5. Figure 1(4) is inconsistent with the text. The figure shows only the PAC pathway, yet all main results come from the full model with the waveform branch and gated fusion. The classification head is drawn as a multi-layer MLP, while Eq. 15 is a single linear layer. The data flow in the transformer block is ambiguous: depending on reading direction, it starts with either MHSA or the FFN. The text claims the filterbank and Hilbert transform share one FFT, but the equations, Algorithm 1, and the figure all describe separate per-band steps. The waveform branch is also not specified.
Questions for authors
  1. Given the physiology, what does PAC measure on ECG? Does the PTB gain survive a surrogate that preserves the power spectrum but destroys phase relations?
  2. How sensitive are the results to window length? How are the edge effects of circular filtering and Hilbert estimation handled?
Reviewer_upBbOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MOSAIC, a transformer-based framework for medical time series classification, and the major innovation is to incorporate phase-amplitude coupling (PAC) into representation learning. The method combines a learnable Gaussian filterbank, a differentiable Hilbert transform, and soft phase-binned amplitude features to capture relationships between oscillations at different frequencies. A cross-band transformer learns representations from these coupling features, with an optional waveform branch to capture information not represented by PAC.

The authors evaluate MOSAIC on three subject-independent EEG and ECG datasets (i.e., Alzheimer's disease, Parkinson's disease, and myocardial infarction). The method achieves higher macro-F1 scores than various published transformer baselines, and the ablation experiments demonstrate the benefits of learned frequency bands and PAC features.

Critical strengths
  1. The paper addresses an important problem in medical time series classification by explicitly modeling phase-amplitude coupling. The proposed differentiable framework allows frequency bands and coupling features to be optimized jointly with the classification objective. This provides a nice connection between established physiological signal analysis and deep learning.

  2. The experimental results are encouraging across three medical datasets. MOSAIC achieves improved classification performance compared with several transformer baselines under subject-independent evaluation. The authors also provide ablation experiments examining the contributions of learned frequency bands, PAC features, the waveform branch, and channel fusion.

Critical weaknesses
  1. The experimental evaluation is limited by the relatively small number of subjects, particularly in the EEG datasets. APAVA contains only 23 subjects, and TDBrain contains 50 subjects. Although the experiments use subject-independent splits, the reported standard deviations are calculated across training seeds using fixed data splits. Therefore, they do not fully capture variability caused by the selection of training and test subjects. Additional evaluation across different subject-level splits would help establish whether the reported improvements generalize to unseen subjects.

  2. The interpretation of learned phase-amplitude coupling (PAC) remains insufficiently validated. Although the method is motivated by physiologically meaningful cross-frequency interactions, the experiments mainly evaluate classification performance. The learned coupling representations may also reflect differences in signal amplitude, spectral power, or other waveform characteristics rather than genuine phase-amplitude coupling. The authors include amplitude-only and phase-alignment controls that are useful. It would also be helpful to have additional validation using simulated data with known coupling patterns to demonstrate that the learned representations accurately capture the known physiological relationships.

Questions for authors
  1. Can the authors evaluate MOSAIC across multiple subject-independent splits or using leave-one-subject-out cross-validation? This would help assess the robustness of the results, especially for APAVA and TDBrain, which have relatively few subjects.

  2. Can the authors provide experiments using synthetic signals with known phase-amplitude coupling to evaluate whether the learned frequency bands and coupling features recover the underlying coupling structure?

  3. How sensitive is MOSAIC to the number of frequency bands, the number of phase bins, and the regularization parameters? Additional sensitivity analysis would help clarify the robustness of the proposed framework.

  4. The baseline results are taken from previously published experiments, while MOSAIC is evaluated separately. Can the authors clarify whether preprocessing, data splits, and model selection procedures are fully consistent across methods? It would also be helpful to provide statistical comparisons to determine whether the reported improvements are significant.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.