acceptodds
Under review as a conference paper at ICLR 2027

ProtoRoute: Dual-Timescale Persistent Evidence Learning for Medical Time-Series Classification

Abstract

Existing medical time-series classification (MedTSC) methods primarily encode class knowledge in network parameters or prototypes optimized on the same gradient timescale as the encoder. We argue that transient structure and persistent class evidence have different statistical scopes and should therefore be modeled and updated separately. Based on this perspective, we propose , a dual-timescale framework that couples a parametric encoder with a -evolving non-parametric class-evidence dictionary. The encoder-agnostic fast pathway captures transient representations of the current sample, while the persistent pathway maintains reusable class evidence across training observations. To read from this persistent state, we introduce hierarchical Token–Class–Prototype Slot routing, where Top- prototype aggregation preserves heterogeneous within-class evidence modes and Top- token aggregation suppresses dilution from weak background evidence. To evolve the persistent state, we further introduce Loss-Adaptive Prototype Evolution, an explicit epoch-level state transition that selectively incorporates reliable and complementary evidence while adapting class-wise replacement to training dynamics. Experiments on 8 datasets demonstrate the strong superiority of ProtoRoute.

Primary area:
learning on time series and dynamical systems
Keywords:
Medical time-series classification, Dual-Timescale, cross-sample evidence

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_EFMXOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ProtoRoute for subject-independent medical time-series classification. It combines a parametric encoder with a non-parametric class-evidence dictionary, using sparse token–class–prototype routing to retrieve evidence and an epoch-level maintenance mechanism to update the dictionary. Experiments across eight benchmarks, together with controlled analyses and subject-level cross-validation, demonstrate improvements in classification performance

Critical strengths

The paper identifies two challenges, within-class heterogeneity and background evidence dilution, and addresses them through multi-prototype representation and Top-q token aggregation, respectively. The experimental evaluation is comprehensive and demonstrates substantial performance improvements across several benchmarks. It includes controlled synthetic studies, matched memory-update comparisons, component ablations, backbone portability experiments, and subject-level cross-validation. This work improves generalization to unseen patients through persistent class-evidence learning. This is clinically relevant, as models must remain reliable despite substantial variation across patients.

Critical weaknesses
  1. The experiments support the effectiveness of the framework, but the explanation of how the two update timescales interact could be clearer. Beyond describing a stability–adaptation balance, could the authors explain how keeping the dictionary fixed within an epoch influences encoder learning, and how selective epoch-level updates contribute to cross-subject generalization? Connecting these mechanisms to the existing results would help clarify the contribution of the update schedule to the overall gains.
Questions for authors

See weakness.

1 comment

Dear Reviewer EFMX,

We are sorry to bother you again. With the discussion period closing in less than 24 hours, we would like to check whether our response and revision have resolved your concerns. We would greatly appreciate any feedback, and we remain happy to answer further questions.

Best regards,

The Authors

Reviewer_hGEuOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes ProtoRoute, a dual-timescale framework for medical time-series classification. The central idea is to separate transient within-sample representations from persistent cross-sample class evidence. Specifically, the framework combines a fast parametric encoder with a slowly evolving non-parametric prototype dictionary. Persistent Evidence Routing (PER) retrieves class-relevant information through hierarchical token-class-prototype matching, while Loss-Adaptive Prototype Evolution (LAPE) updates the dictionary at the epoch level based on candidate reliability, diversity, and training dynamics. Extensive experiments are conducted on 8 medical time-series benchmarks under subject-independent evaluation. The reported results are strong, with ProtoRoute achieving highly competitive performance across almost all datasets and evaluation metrics.

Critical strengths
  1. Interesting motivation and dual-timescale learning design. The motivation behind separating transient representations from persistent class evidence is interesting and reasonable. In medical time-series classification, discriminative information may only appear in certain local segments and may not be consistently present across all samples of the same class. Maintaining reusable class-level evidence while letting the encoder adapt to individual samples offers an interesting perspective on representation learning. The dual-timescale design, in which the encoder and prototype dictionary are updated at different frequencies, is particularly appealing.

  2. Comprehensive experimental evaluation. The experimental section is relatively thorough. The authors evaluate the proposed method on eight datasets covering different modalities and classification tasks. In addition to comparisons with recent baselines, the paper includes component ablations, sensitivity analyses, robustness experiments, synthetic studies, and backbone portability evaluations. These experiments provide useful evidence supporting the effectiveness of the proposed framework.

  3. Strong empirical performance. ProtoRoute achieves impressive results across the evaluated benchmarks, particularly under subject-independent settings. The improvements over existing methods are encouraging and demonstrate the potential of explicitly maintaining persistent class evidence for medical time-series classification.

Critical weaknesses
  1. The overall presentation needs substantial improvement. Despite the interesting motivation and promising results, the manuscript is difficult to follow. In particular, Figure 1 is overly complicated and visually cluttered. It attempts to illustrate too many components and interactions simultaneously, making the overall method pipeline difficult to understand. I strongly recommend reorganizing the illustrations into two separate figures. Place the first in the Introduction to provide a simplified, high-level overview of the proposed framework, highlighting the interaction between the fast encoder and the slow prototype dictionary. The second could be placed in the Method section and illustrate the detailed operations of PER, LAPE, and other important components. The writing and formatting also require careful proofreading. Review terminology, punctuation, and hyphenation conventions for consistency. For example, the expressions "view-wise" and "near–far" around Line 310 use different dash conventions, and similar formatting choices should be checked throughout the manuscript. Overall, the presentation should be considerably more polished to improve readability.

  2. The mathematical formulation is unnecessarily complicated and obscures the main contributions. The Method section introduces a large number of mathematical symbols, definitions, and equations, making it difficult to identify the key methodological ideas. While mathematical rigor is important, many of the formulations describe relatively straightforward operations that could be communicated more effectively through concise explanations. I suggest substantially simplifying the mathematical presentation and focusing on the essential operations that distinguish ProtoRoute from existing approaches. For example, the dual-timescale optimization, prototype retrieval, and dictionary evolution could first be explained through an intuitive step-by-step workflow. Only the most important equations need to remain in the main text, while implementation-specific formulas, detailed update rules, and auxiliary definitions could be moved to the Appendix. The authors already provide pseudocode in the Appendix, which is helpful. A simplified version of the training procedure in the main text would further improve clarity. In its current form, the excessive mathematical notation makes the method appear more complicated than necessary and distracts from its actual contributions.

  3. Additional sensitivity analysis of the loss coefficients is needed. The final objective combines three loss terms: classification loss, prototype loss, and margin loss. Although Table 2 includes ablations that remove the prototype and margin losses, these experiments do not fully characterize the sensitivity of the model to their relative weights. In particular, the coefficients (\lambda_p) and (\lambda_m) introduce additional optimization choices, and performance improvements from multi-loss objectives can depend on their relative magnitudes. I recommend conducting sensitivity analyses for these coefficients, on at least one dataset (e.g., ADFTD), ideally including their interactions, to demonstrate that the reported improvements are robust rather than dependent on carefully selected loss weights. I acknowledge that Appendix C.9 already provides sensitivity analyses for several architectural and memory-related hyperparameters. However, a corresponding analysis of the loss coefficients would provide stronger evidence that the proposed optimization objective is stable.

Questions for authors

N/A

1 comment

The manuscript looks much better than the initial version. I raised my score.

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

This paper proposes ProtoRoute, a dual-timescale framework for medical time-series classification:

  • (1) Persistent Evidence Routing (PER), which performs hierarchical Token–Class–Prototype Slot routing with sparse Top-r prototype and Top-q token aggregation
  • (2) Loss-Adaptive Prototype Evolution (LAPE), which updates the dictionary at epoch boundaries using candidate reliability, class-wise loss dynamics, diversity, and structured prototype replacement.
Critical strengths
  • Good idea root. Time series classification or classification task in general has 2 axes to exploits: within-sample structure and cross-sample class evidence, this intuition is sensible and has been realized / supported by existing general literature. Leveraging both 2 axes allow more reliable and robust signal compared to most works that only focusing on sample-level information.
  • Comprehensive and logical notation sets. This make the mathematical interpretation clear and well support the text main idea, especially for the non-parametric updates of persistent class-evident dictionary.
  • Strong empirical result. ProtoRoute shows its lead in most metrics consistently across 8 benchmarks, with large margin compared to 2nd best baseline(s). The statistical significances are also reported.
  • Comprehensive ablation study. The authors cover several dimensions in ablation: architecture, sparse information routing / aggregation, memory maintenance, and regularization. This provide great insights about how each dimension / contribution affect overall performance: most loss resulted from Architecture ablation and memory maintenance - which is interesting since they are the 2 parts aggregate the most contributions.
Critical weaknesses
  • Figure 1 is hard to digest in its current form. The arrangements of modules in the illustration can be improved: the temporal hierarchy go on top, then come global semantic encoder block, then come the persistent class dictionary (as the global semantic encoder have connection to both, so putting them in middle would make it cleaner and easy to digest). In addition, many texts overlay on graphics, make it hard to read.
  • The contribution of persistent evidence to subject-independent generalization requires more careful interpretation. Although the results are strong, the experiments do not fully establish that the learned dictionary captures subject-independent class evidence rather than additional discriminative information provided by an external memory mechanism. It would be great to have an attempt to interpret the final learned dictionary, categorized by the classes they represent.
Questions for authors

Persistent evidence validity: Can the authors provide further quantitative evidence that the dictionary represents transferable class-level information rather than subject-specific patterns?

Reviewer_4aX5OpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper works on medical time‑series classification for unseen test subjects. The authors build ProtoRoute, which runs two separate parts: a fast‑updating encoder to handle each individual sample’s signal, plus a slow‑changing prototype dictionary that collects reusable class patterns across different patients. They design PER to pull useful information from this dictionary with two sparse selection rules, and LAPE to carefully refresh the dictionary only after each training epoch instead of every batch. They test their method across eight real‑world datasets. Results show their approach beats recent baselines, works well with limited training data, and the core components each bring real performance gains.

Critical strengths

The core conceptual split between fast per‑sample encoding and slow‑evolving persistent cross‑sample class evidence addresses a real pain point in subject‑independent medical time‑series tasks. The paper validates the timescale hypothesis via targeted timescale‑sweep ablations, synthetic controlled simulations, backbone‑portability tests, plus real‑world benchmark results across eight datasets. Multiple orthogonal experiments converge to support the core idea, which strongly increases confidence the improvement is not just a trick of architecture engineering.

The proposed PER‑LAPE persistent pathway is encoder‑agnostic and can be attached to diverse existing time‑series backbones to boost performance, rather than being locked to the authors’ custom encoder. Additional supporting evidence includes strong low‑data performance and robustness against common real‑world corruptions, showing tangible practical value for medical time‑series applications.

Critical weaknesses

All main LAPE and routing hyperparameters (confidence thresholds, replacement weights, diversity weight, r/q slot/token counts) are fixed globally across datasets. While appendix sensitivity studies show acceptable performance over some parameter ranges, there is no formal analysis or principled guidance for setting these values for new, unseen medical time‑series datasets.

Interpretability is a key selling point for prototype‑based medical modelling, but the analysis only provides qualitative case‑study visualisations of a small number of prototype slots. There are no quantitative metrics measuring whether dictionary prototypes genuinely capture clinically meaningful physiological patterns, or quantifying prototype redundancy/corruption during LAPE evolution.

Questions for authors
  1. The LAPE mechanism relies on many hand‑set global hyper‑parameters (confidence threshold, novelty threshold, replacement weights, diversity weight). The sensitivity analysis shows stability within tested ranges, but can you show results on at‑least one fully held‑out medical time‑series dataset where you directly reuse these exact global hyperparameters with zero additional tuning?

  2. In the current submission, evidence for what the persistent dictionary learns comes exclusively from a small number of qualitative example visualisations. Can you provide simple quantitative metrics over the full dictionary (e.g. prototype redundancy across slots, fraction of slots that map to known physiologic patterns, slot contamination by subject‑specific nuisance artefacts) computed on one major benchmark?

  3. When you attach PER+LAPE to external off‑the‑shelf backbones, token representations come from different pre‑existing model feature spaces. Are there concrete failure conditions you have observed where the persistent‑evidence module fails to deliver gains or even hurts performance when plugged into an external encoder? If yes, what properties of the feature space trigger those failures?

  4. In your timescale ablation, epoch‑wise LAPE performs best among tested update frequencies. What intrinsic dataset/model property determines whether epoch‑level updates remain the optimal choice? For example, dataset size, label noise level, feature‑space convergence speed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.