acceptodds
Under review as a conference paper at ICLR 2027

Support-Induced Diagnostic Probing for Multichannel Medical Time Series

Abstract

Multi-channel medical time-series classification requires combining diagnostic context (usually given by labelled examples) with the information contained in a new recording. We propose Support-Induced Diagnostic Probing (SIDP), a class-conditioned retrieval architecture that converts positive-vs-negative support differences into diagnostic probes. Cross-channel attention first transforms a recording into a fixed-size set of latent representations. Support-generated probes then retrieve and score diagnosis-specific information from these representations. Once trained on a dataset, SIDP classifies new recordings from a small labelled support context without retraining or updating its parameters. We evaluate on PTB-XL, a large public ECG benchmark, where the task is to predict five diagnostic categories that can co-occur within a recording. Using the standard patient-disjoint test split, SIDP achieves **0.8450** macro-AUROC and **0.6215** macro-F1 across 2,158 test records from 1,877 patients. Our method improves macro-F1 over pooled-descriptor MLP and support-conditioned MLP baselines by **7.50** and **9.97** percentage points, respectively, with bootstrap confidence intervals that exclude zero. We further perform ablation experiments on the PTB dataset to isolate architectural contributions when identical channel descriptors are provided. Here, SIDP reaches **0.8603** window macro-F1 and **0.9661** patient macro-F1. Cross-channel attention improves window macro-F1 by **14.13** points over a pooling baseline with the same inputs, and support-generated probes improve macro-F1 by **2.14** points over fixed learned queries. An analysis of the attention weights suggests that both attention stages adapt to each recording. These results suggest that support-induced probing is an effective way to retrieve diagnostic information from multi-channel recordings. The source code for our architecture is available at https://anonymous.4open.science/r/sidp-85DD

Primary area:
learning on time series and dynamical systems
Keywords:
time-series classification, support-conditioned learning, in-context learning, few-shot classification, class-conditioned retrieval, multichannel time series, medical time series, EEG, ECG

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 2 reviews, above 27% of the others.
Reviewer_GubjOpenReview ↗
Rating
2
Confidence
3
Informativeness
3(+1)
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SIDP, a method for classifying multichannel physiological signals such as EEG and ECG. It uses a small set of labeled support examples to build class-specific probes, and these probes extract diagnostic information from a new recording. The support set is used only at inference, so the model does not need retraining. The method is evaluated on EEG and ECG datasets in both single-label and multi-label settings.

Critical strengths
  1. All notation, tensor shapes, and hyperparameters are clearly defined, which makes the work highly reproducible. The notation is consistent and carefully organized, and it greatly helped me understand the core model architecture.

  2. The experimental results are reported in a detailed and transparent way. The authors present both the cases where the model performs well and its limitations, so the reader can clearly see the strengths and limits of the proposed method.

  3. The ablation studies are systematic. They make it possible to see how each component of the architecture contributes to the final performance.

Critical weaknesses
  1. Insufficient baseline comparisons. I understand the authors’ point that the input representations differ across methods. Even so, many models could be compared from the perspective of the same classification task, regardless of input format. Direct comparisons with such models, or experiments with re-implemented models of similar architecture, seem necessary.

  2. The performance gains appear to come from the input representation rather than the architecture. As the authors acknowledge, whether the margin input is used appears to have a large effect on performance. A more fundamental concern is the large gap in parameter count between SIDP and the MLP baseline. This raises questions about the efficiency and necessity of the proposed architecture. Because the models are not matched in size, the observed performance differences are hard to attribute to architectural advantages.

Questions for authors
  1. Figure 1 has low readability, and some of its text appears to overlap, so it should be revised. It is also placed after the Method section, which makes it hard to read alongside the corresponding text.
  2. From a classification-task perspective, there appear to be models whose performance could be compared regardless of input format. Could the authors provide such comparisons, or explain why they were not included?
  3. What is the practical benefit of using a support set? The performance gain from support-induced probes does not appear substantial. Could the authors elaborate on this point?
  4. The benefit of the proposed architecture appears to be significant mainly on the PTB dataset. Are there structural advantages that would let it generalize to other datasets?
  5. Is the per-recording normalization in Section 4.5 appropriate? Since it uses statistics from each test patient’s own recording, could it cause information leakage?
3 comments

We thank the reviewer for the positive comments on clarity and reproducibility. Locations refer to the revised PDF.

Q1. Readability and placement of Figure 1

We agree. Figure 1 has been redrawn at its printed size, so all labels are now about 7–8 pt and no labels overlap. The figure now sits at the start of the Method section (page 4, beside Secs. 3.1–3.2). Each block names the equation it implements.

Q2. Task-level comparisons beyond the descriptor input

We agree that the complete system should also be evaluated at the classification-task level. Table 2 (Table 3 in the submission) is now the primary overview of related methods. We also added task-level comparisons in which other models receive SIDP's input: an MLP, a wide MLP, a Transformer, a Set Transformer, a support MLP, and ProtoNet and Matching Networks with the SIDP encoder architecture (Table 1, lower block). On these identical inputs, SIDP scores above the attention-based and support-conditioned models, and the flattened MLP scores slightly above SIDP.

Q3. Practical benefit of labelled supports

The practical capability is that the diagnostic context can be changed by supplying labelled examples, without updating the trained model's parameters. Summaries of the positive versus negative supports generate the probes, and these probes determine what the retrieval from the query attends to. This capability does not, by itself, guarantee better accuracy than a fixed-label head.

Multi-label PTB-XL. The support-source experiment (Sec. 4.4, Table 3) gives direct evidence that the benefit can extend to support patients who were outside encoder training. SIDP's F1 difference from fixed queries is +0.85 [+0.12, +1.50] points with training supports and +0.72 [+0.05, +1.38] with validation-patient supports. With test cross-split supports it is +0.64 [−0.05, +1.22]: the point estimate is positive, but this interval does not establish an advantage.

Original PTB ablation. With training-patient supports, SIDP exceeds fixed learned queries by +2.14 [+0.75, +3.61] F1 points. With validation-patient and test cross-split supports, the differences are +2.07 [+0.75, +3.47] and +2.06 [+0.74, +3.44]. Under nested patient cross-validation, the estimated difference from fixed queries is +0.51 [−0.38, +1.41].

Other comparisons. On single-label PTB-XL with margin inputs, SIDP improves on the support MLP by +2.04 [+1.21, +2.88] F1 points after correction for six comparisons, while the ordinary MLP performs better.

Q4. Generalization beyond PTB and structural advantage

The structural capability is this: a target descriptor derived from the supports controls what is retrieved from the query's cross-channel latent slots. This connects the complete input-processing pipeline to evidence selection conditioned on the supports. Evaluating only fixed-label accuracy would miss this functionality, and evaluating only against pooled inputs would leave unclear where the accuracy gains come from. We therefore now report both system-level and matched-input evidence.

Generalization remains conditional:

  • On multi-label PTB-XL, SIDP has higher AUROC and F1 point estimates than the tested attention-based and support-conditioned comparators. Its F1 advantage over fixed queries persists with validation-patient supports.
  • The ADFTD and single-label per-channel PTB-XL comparisons also favor the MLP (new Sec. 4.7).

Q5. Per-recording normalization and information leakage

Per-recording normalization is appropriate for an offline protocol: one where the complete query recording is available and its statistics are computed from it alone, without labels, other patients or later recordings. If a decision is made per window or online, using the entire recording changes the information available at prediction time. For multi-label PTB-XL, the statistics use only the ten-second record being classified.

Per-channel SIDP macro-F1 under alternative normalization methods (Table 39):

DatasetPatient/recording statisticsWindow-only statisticsTraining-only statistics
PTB0.82020.80910.7916
ADFTD0.44210.42850.4121

The table has no paired intervals for these within-SIDP changes, so it does not establish that the changes are statistically significant, nor that SIDP is invariant to normalization.

The margin-input sensitivity analysis (Table 26, Table 25 in the submission) is a different comparison. Its intervals for the window-only and training-only changes include zero. The submission's sentence "we conclude that there is no clear drop" has been removed.

The ethics statement now notes that per-window or online use requires window-only statistics. For the window-level PTB and ADFTD analyses, we will make window-only normalization the default in the final version and also report the recording-level and training-only alternatives.

The readability of the figure has improved. The authors also appear to have substantially strengthened the baseline comparisons under identical input conditions. I also appreciate that they added a comparison with a parameter-matched MLP, which directly addresses my concern about the difference in model size. However, I still have concerns about the fact that SIDP performs significantly worse than a simple MLP of the same size (Table 2). The authors explained the cause, but this seems to imply that the advantage of the proposed architecture does not show up in terms of accuracy.

In addition, SIDP’s advantage still does not hold on datasets other than multi-label PTB-XL (ADFTD and single-label PTB-XL), which remains a major weakness. This is because it raises the risk that the benefits of the proposed method depend on a specific dataset or data split. Overall, the authors have responded sincerely to the questions I raised.

Therefore, I have raised my Informativeness score to 3. However, I will maintain my overall rating of 2. Please let me know if I have misunderstood anything.

We thank you for the follow-up and for raising the Informativeness score. Your reading of Table 2 is correct, and we address both concerns below:

1. Accuracy against the equal-size MLP

You are right that the gap is significant. Under nested cross-validation, SIDP reaches 0.8604 AUROC and the equal-size MLP reaches 0.8705, a difference of 1.01 [0.77, 1.23] points. We agree that, on the five training labels, the advantage of SIDP is not higher accuracy. We do NOT claim it in the paper.

The advantage is a capability that the MLP does not have. The MLP learns a separate output weight vector for each of its five labels. A new diagnosis, therefore, needs new labelled data and retraining. SIDP has no label-specific parameters. It builds the classifier for each label at inference from ten labelled patients. In every fold of the same nested experiment, the trained SIDP scored 30 PTB-XL statements that were not among its training labels, with a mean AUROC of 0.715 and no retraining. The MLP cannot produce a score for these statements at all without retraining.

The one-point gap is the price of this capability. Fixed learned queries keep the SIDP encoder but learn one set of queries per label, like the MLP. With learned queries, the encoder comes within 0.20 AUROC points of the MLP. Most of the gap, therefore, comes from building each label's classifier from ten examples instead of learning it from the full training set. It does NOT come from a weaker encoder. On identical inputs, SIDP is also the strongest of the models that work from labelled examples. It scores above ProtoNet and Matching Networks with the same encoder, and above the Transformer and Set Transformer.

2. Results on ADFTD and single-label PTB-XL

We agree that this is a limitation, and the paper reports it in Section 4.8. Following two details may help to judge how specific it is to SIDP:

  • On ADFTD (52 training patients), the MLP is ahead of every attention-based model, not only SIDP. Fixed queries, Transformer, Set Transformer and SIDP reach 0.4105, 0.4100, 0.4032 and 0.4137, against 0.5257 for the MLP. SIDP has the highest score of these four. The gap, therefore, appears to be shared by attention-based models on this small cohort, rather than specific to the probe mechanism.
  • On single-label PTB-XL with per-channel inputs, the low score of SIDP points to training the encoder from a random initialization. In diagnostic runs where the encoder is warm-started, SIDP reaches 0.3999, level with fixed queries at 0.3911 (+0.88 [-0.25, +1.95]). With the margin input on the same dataset, SIDP is 2.04 [1.21, 2.88] points above the support MLP, while the MLP without supports is 0.96 points higher.

On the risk of dependence on one split, the multi-label PTB-XL results in Table 2 come from ten-fold nested patient cross-validation, NOT from a single split. The ordering of SIDP and the equal-size MLP is the same in all ten folds. On PTB, SIDP has the highest mean of the seven methods on the standard split, although its gains over the MLPs are not established under nested folds.

We hope this clarifies that SIDP is meant to trade a small amount of accuracy on fixed labels for labels that can be defined at inference. We are happy to answer any further questions.

Reviewer_qE1WOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Support-Induced Diagnostic Probing, a class-conditioned retrieval architecture for multichannel medical time series. Cross-channel attention first maps each recording into a fixed-size set of latent slots, which the later probes attend to. Positive and negative support means then form a class descriptor that offsets learned probe seeds, and a shared scorer scores every class. The paper frames the probes as a way to classify new recordings from labelled support examples without retraining the encoder. The authors evaluate the method on three public datasets with patient-disjoint splits, and on the standard PTB-XL multi-label split it reaches 0.8450 macro-AUROC and 0.6215 macro-F1. The paper also reports matched-input comparisons in which the advantage over simple baselines is small or absent, and it says so in the appendix.

Critical strengths
  1. Cohort Discipline: The multi-label split states every exclusion rather than dropping records silently, and the AUROC gain holds at 6.31 points over the MLP on the test patients with no single-label overlap. That subset is the cleanest available test of whether the development cohort drives the reported result.
  2. Reported Nulls: The paper states that an MLP on the same margin input also outscores every reference entry, and it concludes that this gain mainly reflects the margin input. It also reports the MLP without support scoring 0.96 points above SIDP on single-label PTB-XL, which most authors would have left out.
  3. Controlled Ablation: All seven methods receive the same per-channel PTB tensor and the same 32 support sets, and the learned methods share three seeds and one schedule. Each of the three controls changes a single component, so the probe-level differences on PTB are measured rather than confounded with input.
Critical weaknesses
  1. Headline Gain Confounds Input and Size: The multi-label comparison gives SIDP segment descriptors and gives every baseline pooled descriptors averaged over leads and segments. That choice makes the headline gain a comparison of inputs and classifiers together, which the paper itself concedes by calling it a combined effect. No segment-descriptor MLP appears anywhere, and the authors note that they ran no control with equal input information. SIDP also uses a much larger model than the baselines it beats, so the 7.50-point gain over the MLP in Table 2 cannot be assigned to probing. A reader therefore cannot tell how much of the gain comes from the token structure and how much from the probes.
  2. Mechanism Collapses on Per-Channel PTB-XL: On single-label PTB-XL with per-channel inputs, SIDP reaches 0.1540 window macro-F1, below the fixed-query control at 0.3911 and below a plain MLP. The abstract instead reports a gain over fixed queries from the per-channel PTB ablation, so the advantage does not carry across datasets. Falling below a nearest-centroid prototype is easier to read as a fault in the narrower input path than as a transfer failure. Whatever the cause, the paper offers no evidence that the probe mechanism generalizes beyond PTB.
  3. Deployment Claim Remains Untested: The abstract promises classification of new recordings from a small labelled support context, but every support window comes from a training patient. No reported experiment draws support from held-out subjects, so each probe is built from a cohort the encoder has already memorised. Appendix H adds that the single-label margin pipeline was selected because it scored above all twenty-two reference entries on test scores, a choice the authors disclose. The multi-label headline uses a separate pipeline, so that exposure is confined to the single-label results, though the claim in the abstract covers both.
Questions for authors
  1. Suggestion: Run both MLPs on the same segment tensor SIDP receives, since that control alone would decide whether the headline gain belongs to the classifier or the input.
  2. Why does SIDP fall to 0.1540 on per-channel PTB-XL while fixed learned queries reach 0.3911, and does that path share latentizer and probe code with the PTB path?
  3. Can the evaluation support sets be rebuilt from patients outside the training split, with SIDP compared against fixed learned queries under that protocol?
  4. Would the reported pipeline survive validation-only selection, given that Appendix H records a higher validation score for the anchored candidate it rejected?
4 comments

We thank the reviewer for the careful reading and for recognizing the cohort discipline, the reported negative results and the controlled ablation.

Q1. Equal-input MLP controls and attribution of the headline gain

Keeping lead and segment information separate is part of the pipeline we evaluate. The comparison with pooled-descriptor systems therefore measures that combined design choice, and it does not isolate the benefit of the diagnostic probes.

Both MLPs receive exactly the same 12 × 4 × 50 segment tensor that SIDP receives. They use the official PTB-XL folds and the same training schedule, and their checkpoints and thresholds are selected on validation fold 9 only.

Multi-label PTB-XL, identical segment tensorMacro-AUROCMacro-F1
Flattened MLP, 64–640.85140.6273
Wide MLP, 256–2560.84830.6187
Flattened linear model0.84010.6168
Flattened support-conditioned MLP0.83470.5995
SIDP0.84500.6215

With the structured input, the MLP reaches 0.6273 F1, compared with 0.5465 for the MLP with pooled input: an 8.08-point difference. The structured input is therefore a substantial source of the full-pipeline improvement. On these matched inputs, SIDP is 0.64 [0.07, 1.20] AUROC points below the flattened MLP in the paired analysis.

Changes in the revision.

  • Sec. 4.2 describes the 7.50-point gain as a result of the combined pipeline, not of probing specifically. It also states the parameter counts (82,946 for SIDP vs. 7,749 for the pooled MLP).

Q2. The single-label PTB-XL collapse and shared implementation

The PTB, PTB-XL and ADFTD experiments use the same source code for the latentizer and the probes. Shared code, however, does not mean identical optimization behavior across datasets or input representations.

The diagnostics fit two explanations: support contrasts that differ only weakly between classes, and an optimization bottleneck.

  • On five-class per-channel PTB-XL, SIDP's loss stays close to the value for uniform predictions, ln(5) = 1.609, and ends at 1.599 after 1,800 updates. Fixed queries reach 1.317.
  • The SD of the logits across classes is 0.08 for SIDP, versus 0.95 for fixed queries.
  • The SD of the support contrasts is 0.15 here, versus 0.97 on PTB.

A diagnostic run with a warm-started encoder reaches 0.3748 F1, which suggests that initialization and representation learning matter. That run uses a different initialization and is not a reported result.

The reported SIDP result remains 0.1540, compared with 0.3911 for fixed queries. The original five-class result does not support a universal advantage for the mechanism.

Q3. Supports from patients outside encoder training

Every trained model stays fixed, and only the supports change:

  1. Training patients.
  2. Validation patients: 32 draws of five positive and five negative fold-9 patients per label.
  3. A patient-disjoint test cross-split: the test patients are split into two halves, and each half supplies labelled supports for the queries in the other half.

SIDP minus fixed learned queries, in percentage points (positive values favor SIDP; Table 3 in the revision):

MetricTraining supportsValidation supportsTest cross-split supports
AUROC (PTB-XL ML)+0.25 [+0.05, +0.48]+0.18 [0.00, +0.41]+0.15 [−0.02, +0.35]
F1 (PTB-XL ML)+0.85 [+0.12, +1.50]+0.72 [+0.05, +1.38]+0.64 [−0.05, +1.22]
F1 (PTB, per-channel)+2.14 [+0.75, +3.61]+2.07 [+0.75, +3.47]+2.06 [+0.74, +3.44]

On multi-label PTB-XL, the F1 advantage persists when the supports come from validation patients. The encoder therefore does not need to have been trained on the support patients for an F1 advantage to appear.

On per-channel PTB, the difference is almost unchanged across the three sources.

The cross-split protocol still assumes labelled supports from the target population. It does not test the setting where no labelled target patients are available.

Q4. Validation-only selection and Appendix H

The concern is valid. In Appendix H, the pipeline selection was informed by test results.

With validation-only selection, the alternative Appendix H candidate is preferred (mean validation F1 0.7019 vs. 0.6866). That candidate scores 0.8437 on PTB and 0.3492 on ADFTD, compared with 0.8410 and 0.5043 for the pipeline we originally selected. We present this as a selection-sensitivity analysis of two distinct candidate pipelines, not as interchangeable results for one unchanged pipeline.

We also ran nested patient cross-validation of the per-channel PTB ablation, with five outer and three inner folds.

Nested PTB cross-validationDifference in macro-F1
SIDP − canonical mean pooling+5.89 [+2.93, +8.83]
SIDP − fixed learned queries+0.51 [−0.38, +1.41]
SIDP − probe mean pooling−0.24 [−0.87, +0.40]
SIDP − MLP−2.43 [−5.62, +0.79]
SIDP − support MLP−3.11 [−6.05, +0.08]

SIDP's difference from canonical mean pooling remains positive.

Thank you for a revision that answers the central objection with data. You built the equal-input control, put it in Table 1, and then wrote in Section 4.2 that the remaining differences do not isolate the probes. That candour is why this is a list of requests rather than a recommendation to reject.

What is settled

  • The input confound is closed: every method receives the same tensor under the same selection protocol (Table 1).
  • The consequence is stated rather than left to the reader (Section 4.2).
  • Code sharing is checked more strongly than I asked for (Appendix K).
  • The support-source experiments are the strongest evidence in the revision for the mechanism (Table 3).

What needs action

  1. The comparison now works against the paper. On identical inputs a flattened MLP is ahead on both metrics and a wider one is ahead as well (Table 1). Neither row carries a parameter count, and my own lower bound for the smaller model already exceeds the whole method. Please give counts for every row and say what the probes still buy that a fixed-label model cannot.

  2. The abstract. It reports the flattened MLP's score but leaves it out of the comparators the method beats, and the wider model is absent entirely. Please reconcile the list with the table.

  3. Fixed learned queries has no absolute row in the main block, so the decomposition requires subtracting one table from another. Please give it a row and a count.

  4. The per-channel reversal is disclosed, not repaired. The diagnosis is careful and the conclusion is one I would have written, but no setting recovers the result. Please report one setting in which it does, even as a diagnostic.

  5. Validation-only selection undercuts one headline. The candidate chosen on validation alone falls below the reported pipeline and below entries in the comparison table (Section 4.3). Please repair or restate that claim.

  6. The decisive evidence is narrow. It is a single metric on two datasets, and neither advantage holds its interval under the harder protocol. Please say which single result the paper wants to be judged on, and make the title and abstract claim that and no more.

Where this leaves me

I keep my assessment at marginally below the acceptance threshold, and would not mind seeing it accepted. The audit is good and the reporting honest, which is why this is not a reject, but the contribution is thinner than the title promises. One comparison under nested folds, or on a cohort nobody has inspected, in which the method matches the flattened MLP at equal parameter count while keeping its probes swappable, would move it up a band. I would rather see this paper fixed than rejected.

AuthorsOpenReview ↗

Reply to Reviewer qE1W (1/2)

We thank you for your time and effort. Before the individual points, we explain why SIDP is useful even though a fixed-label MLP scores slightly higher on the training labels. In clinical practice, the set of diagnoses of interest is not fixed. New diagnostic statements, finer subtypes and site-specific criteria appear after a model is deployed. A fixed-label model, such as the MLP, needs labelled data and retraining for every new label. SIDP has no label-specific parameters, so the same trained model can score a new label from a few labelled patients without retraining. SIDP offers this flexibility while staying within about one AUROC point of an MLP of equal size on the training labels, and above the Transformer, Set Transformer, ProtoNet and Matching Networks on identical inputs. We ran the experiments as you suggested and address each point below. The new experiments follow an analysis plan whose hash was stored before any model was trained. Intervals are paired 95% patient-bootstrap intervals and the differences are in percentage points.

Due to the 5,000-character limit, our response is split into two comments: this comment addresses points 1-5, and the next comment addresses points 6 and 7.

1. Parameter counts:

Identical 12×4×50 inputParameters
Flattened linear12,005
Flattened MLP, 64-64158,149
Wide flattened MLP, 256-256681,733
Flattened support MLP465,089
Transformer76,101
Set Transformer177,413
ProtoNet, SIDP encoder37,378
Matching Networks, SIDP encoder37,378
Fixed learned queries, SIDP encoder84,226
SIDP82,946

The pooled baselines have 255 (linear), 7,749 (MLP), 13,889 (support MLP) and 0 (prototype) parameters. We would update every row of Table 1 to carry its count. SIDP is among the smaller models. The 64-64 MLP has 1.9 times and the wide MLP 8.2 times as many parameters.

What the probes buy is a label set that is defined at inference rather than fixed by the trained weights. SIDP has no label-specific parameters. In every outer fold of the nested experiment below, the same trained SIDP scored 30 PTB-XL statements outside its five training labels, using five positive and five negative patients per statement as supports and no parameter update. It reached a mean AUROC of 0.715 over 300 fold-statement pairs. We will add this analysis to the revised manuscript. A fixed-label model, including the MLP, cannot output a dedicated score for these labels without retraining.

2. The abstract: We would update the manuscript to list every identical-input comparator with the direction of each difference in the abstract. On the standard split, SIDP is above seven (the linear model, fixed learned queries, ProtoNet, Matching Networks, Transformer, Set Transformer and the support MLP) of the nine comparators on both metrics. Only the two flattened MLPs, which are 1.9 and 8.2 times larger, are ahead, with the 64-64 MLP on both metrics and the wide MLP on AUROC only. The abstract would also report the nested equal-parameter result (point 7).

3. Fixed learned queries: We would update the table 1 to have the row: 0.8425 AUROC and 0.6130 macro-F1, with 84,226 parameters. On the standard split, SIDP (0.8450 and 0.6215) is above it by +0.25 [+0.05, +0.48] AUROC and +0.85 [+0.12, +1.50] macro-F1 points, with 1,280 fewer parameters.

4. A setting in which the per-channel result recovers: These are diagnostic runs on single-label PTB-XL with per-channel inputs, 3 seeds each. We will add them to the single-label PTB-XL analysis (Section 4.7 and Appendix K.4).

SettingMacro-F1
SIDP, 1,800 updates (reported)0.1540
SIDP, 5,400 updates0.2245
SIDP, encoder warm-started from the fixed-query fit, 5,400 updates0.3999
Fixed learned queries (reported)0.3911

With a warm-started encoder, SIDP is level with the reported fixed-query result (+0.88 [−0.25, +1.95]). Once the encoder is trained, the support-generated probes perform as well as learned queries on this dataset, so the reported score points to optimization from a random initialization rather than to the probe mechanism.

5. Validation-only selection: We will restate the claim in Section 4.3. Since an MLP on the same margin input also exceeds all 22 reference entries, we attribute that gain to the margin input and will present the reference comparison as context, in the appendix. The pipeline chosen on validation alone scores 0.8437 on PTB, above the reported pipeline, and 0.3492 on ADFTD, below five of the fourteen ADFTD entries.

AuthorsOpenReview ↗

Reply to Reviewer qE1W (2/2)

6. The result we ask the paper to be judged on: SIDP is designed to define a diagnostic label at inference from a few labelled patients, with no label-specific parameters. Even so, it remains competitive with fixed-label models that learn a dedicated output for each label. On PTB-XL, under nested patient cross-validation, it stays within about one AUROC point of an MLP of equal size (0.8604 vs. 0.8705), and on the standard split it scores above the Transformer and Set Transformer.

7. The comparison you proposed: following are to be added to the revised manuscript as a new subsection, table and appendix section.

  • Folds: Multi-label PTB-XL, all ten patient-disjoint folds. Each fold is tested once, training uses eight folds, and the epoch and per-label thresholds are selected on the remaining fold.
  • Models: SIDP is the submitted architecture, which has no label-specific parameters, so its probes are swappable. The MLP's hidden width is reduced to 34 to match SIDP's parameter count.
10 nested foldsParametersAUROCMacro-F1
SIDP82,9460.8604 [0.8571, 0.8637]0.6326 [0.6272, 0.6376]
Flattened MLP, equal size82,9990.8705 [0.8673, 0.8735]0.6478 [0.6420, 0.6529]
Flattened MLP, 64-64158,1490.8721 [0.8689, 0.8750]0.6520 [0.6464, 0.6571]
Fixed learned queries84,2260.8685 [0.8653, 0.8718]0.6424 [0.6366, 0.6479]

At equal size, the MLP is ahead by 1.01 [0.77, 1.23] AUROC and 1.52 [1.08, 1.94] macro-F1 points. The revised manuscript will report this directly.

On the five training labels, the equal-size MLP has an advantage. Both models are trained end to end on the same labels and folds. They differ in how each label's classifier is obtained. The MLP learns an output weight vector for each label from every training patient with that label, whereas SIDP computes each label's probe at inference from only ten support patients. Fixed learned queries isolate this difference, because they keep SIDP's encoder and replace the support-generated probes with learned per-label queries. With learned queries, SIDP's encoder comes within 0.20 AUROC points of the equal-size MLP (point estimate), so the encoder itself is competitive. The remaining difference, 0.81 [0.62, 0.98] AUROC points, is the price of computing each label's classifier from ten examples instead of learning it from the full training set. For that price, SIDP gains the swappable label set in point 1, which no fixed-label model offers.

Comments on the paper

3 comments

We thank the reviewer for the clear questions. Locations refer to the revised PDF.

Q1. Main comparisons on the same structured descriptors

SIDP's contribution is the complete pipeline: keeping channel and segment structure, cross-channel latentization, and support-induced retrieval. The pooled comparison measures the combined effect of these parts. To attribute the gain to its sources, we added controls that receive the same input as SIDP. These are now the lower block of Table 1:

Multi-label PTB-XL, identical 12 × 4 × 50 inputsMacro-AUROCMacro-F1
Flattened MLP0.85140.6273
Wide MLP0.84830.6187
Transformer0.82370.6005
Set Transformer0.83500.6102
Support MLP0.83470.5995
ProtoNet, SIDP encoder architecture0.83720.6118
Matching Networks, SIDP encoder architecture0.82140.5851
SIDP0.84500.6215

On identical inputs:

  • SIDP has higher AUROC and F1 point estimates than Transformer, Set Transformer, Support MLP, ProtoNet and Matching Networks.
  • The flattened MLP scores higher than SIDP on both metrics.
  • The wide MLP has higher AUROC than SIDP, while SIDP has higher F1.

These results show where the headline gain comes from, but they do not make the full-pipeline comparison irrelevant.

Q2. ADFTD, single-label PTB-XL, and when supports help

The two negative results show different behaviors, so we discuss them separately. Both are now in the main text (new Sec. 4.7).

ADFTD. In the comparator analysis, with 52 training patients, fixed queries, Transformer, Set Transformer and SIDP reach 0.4105, 0.4100, 0.4032 and 0.4137 F1, respectively. All of these are below the MLP's 0.5257, so the attention controls also perform poorly. This pattern fits a difficult small-cohort regime, but it does not establish that data scarcity causes the gap.

Single-label per-channel PTB-XL. Here SIDP makes nearly uniform predictions and its loss decreases very little. The loss traces, support-contrast diagnostics and the encoder warm-start experiment support an optimization-bottleneck explanation (App. K.4).

When supports help. The evidence that support conditioning is useful is task-specific (F1 differences, in points):

  • Multi-label PTB-XL: SIDP exceeds fixed queries by +0.85 [+0.12, +1.50] with training supports and by +0.72 [+0.05, +1.38] with validation-patient supports. The test cross-split estimate is also positive, +0.64 [−0.05, +1.22], but its interval includes zero.
  • Per-channel PTB, standard split: the differences from fixed queries are +2.14 [+0.75, +3.61], +2.07 [+0.75, +3.47] and +2.06 [+0.74, +3.44] with training, validation and test cross-split supports, respectively.
  • Single-label PTB-XL with the margin input: SIDP exceeds the support MLP by +2.04 [+1.21, +2.88] after correction for six comparisons. The MLP without supports remains ahead, by 0.96 [0.30, 1.63].

Q3. Generalization after test-informed selection

We acknowledge that the test-informed selection disclosed in Appendix H limits the strength of the evidence from that comparison. We therefore added validation-only selection for the new comparisons, and nested patient cross-validation to check robustness across patient partitions. The multi-label PTB-XL results use a separate pipeline whose checkpoints and thresholds were selected on validation.

Nested cross-validation. On PTB, with five outer folds and three inner folds, fitting and selection were restricted to each outer training partition. SIDP improved macro-F1 over canonical mean pooling by 5.89 percentage points [2.93, 8.83]. Its estimated difference from fixed learned queries was +0.51 points [−0.38, +1.41]. This interval includes zero, so it does not establish superiority across patient splits (Sec. 4.5, Table 37).

These experiments strengthen the robustness analysis, but they cannot make an already-inspected test cohort untouched after the fact. The revised discussion distinguishes these robustness findings from independent confirmation on a cohort that has never been inspected, and lists such a cohort as future work.

Q4. Stronger attention-based and support-conditioned comparators

This comparison is directly aligned with our architectural goal. The reference overview (Table 2, Table 3 in the submission) now comes first in Sec. 4.3, and the controlled attention and support-conditioned comparisons sit next to it in Table 1.

Model size. Table 35 (Table 34 in the submission) reports 82,946 trainable parameters for SIDP, compared with 7,749 for the original pooled-input MLP. The additional attention baselines have 76.1k (Transformer) and 177.4k (Set Transformer) parameters. These are stronger architectural comparators, but they are not an exact parameter-matched MLP experiment. We removed efficiency and architectural-superiority claims that these controls do not support.

AuthorsOpenReview ↗

Reply to Reviewer taFB (1/2)

Dear Reviewer,

As the author-reviewer discussion period is about to end, we hope that our earlier responses have answered your questions. Below, we summarise how the revised paper addresses each weakness, and we add a comparison with published methods.

We first restate why SIDP is useful. Diagnostic label sets change after a model is deployed, as new statements, subtypes and site-specific criteria appear. A fixed-label classifier needs new labelled data and retraining for each of them. SIDP has no label-specific parameters. It defines each label at inference from a few labelled patients, without retraining.

W1. Unfair comparison in the main experiments

  • Every comparator in the lower block of Table 1 receives the identical 12x4x50 segment tensor, and every row reports its parameter count. SIDP scores above seven (the linear model, fixed learned queries, ProtoNet, Matching Networks, Transformer, Set Transformer and the support MLP) of the nine comparators on both AUROC and F1. Only two flattened MLPs, with 1.9 and 8.2 times as many parameters, are ahead on AUROC.
  • Table 2 compares SIDP with an MLP of the same size under ten-fold nested patient cross-validation. The MLP is ahead by 1.01 AUROC points [0.77, 1.23]. With learned per-label queries, the SIDP encoder comes within 0.20 points of this MLP. The gap, therefore, comes from building each label's classifier from ten examples at inference, NOT from a weaker encoder. In return, the same trained SIDP scores 30 PTB-XL statements outside its training labels (mean AUROC 0.715) without retraining. The MLP cannot do this.

W2. Inconsistent performance across datasets

  • ADFTD: Subject-independent ADFTD is hard for attention-based models in published work as well. Under the Medformer protocol (Medformer, Y. Wang et al., NeurIPS 2024), the Transformer reaches 48.09 F1, iTransformer 46.79, PatchTST 41.97 and Medformer 50.65. In our matched analysis, every attention-based model falls below the MLP (52.57). Fixed queries, Transformer, Set Transformer and SIDP reach 41.05, 41.00, 40.32 and 41.37, and SIDP is the highest of the four. The gap, therefore, appears to reflect the small cohort (52 training patients) rather than the probe mechanism.
  • Single-label PTB-XL: The low per-channel score points to training the encoder from a random initialization. In diagnostic runs with a warm-started encoder, SIDP reaches 39.99, level with fixed queries at 39.11 (Section 4.8, Appendix K.6).
  • Multi-label PTB-XL: The ordering of SIDP and the equal-size MLP is the same in all ten nested folds. The main results, therefore, do not depend on one split.

W3. Test-informed pipeline selection

  • The issue in Appendix H concerns only the single-label margin pipeline. The multi-label PTB-XL pipeline, which carries the main claims, selected checkpoints and thresholds on validation data only.
  • All comparisons added in the revision followed an analysis plan fixed before training, with validation-only selection.
  • The ten-fold nested cross-validation on PTB-XL (Table 2) and the nested cross-validation on PTB use no test data for any choice.
  • Validation-only selection of the single-label pipeline gives 0.8437 on PTB, above the reported 0.8410, and 0.3492 on ADFTD. We report both, and we no longer claim that SIDP outperforms the reference methods (Section 4.4).
AuthorsOpenReview ↗

Reply to Reviewer taFB (2/2)

Q4. Comparison with published attention-based and support-conditioned methods

PTB (MI vs healthy controls, subject-independent 60/20/20 split). Published scores are as reported in each paper.

MethodF1 (%)
FORMED, pretrained foundation backbone (Huang et al., ICLR 2026)92.35
SIDP, per-channel input (ours)86.03
ViRe (Yu et al., NeurIPS 2026)85.81
HM-BiTCN with CIF (Hu et al., arXiv 2025)85.59
MLP on the same input as SIDP (ours)84.13
TeCh (Yu et al., ICLR 2026)81.97
MedGNN (Fan et al., WWW 2025)80.40
Medformer (Y. Wang et al., NeurIPS 2024)79.18
iTransformer (as reported in Medformer)79.06
Transformer (as reported in Medformer)68.47
PatchTST (as reported in Medformer)64.36

SIDP scores above every published model in this table except FORMED, which repurposes the pretrained TimesFM forecasting model. These comparisons are unpaired. The published models use 1.2-second samples (300 time steps at 250 Hz) from 198 subjects, while we use 1-second windows from all 200 patients with the same subject split ratio. Our PTB test patients were also inspected during development (Appendix H).

New labels without retraining: We found no published method that scores held-out PTB-XL statements from a support set. The closest are zero-shot ECG-text models pretrained on about 800,000 MIMIC-IV ECGs. KED reports a mean AUROC of 0.744 over 46 PTB-XL statements (Tian et al., Cell Rep. Med. 2024). On the PTB-XL superclasses, D-BETA reports 0.762, and 0.742 for MERL in the same table (Pham et al., ICML 2025). SIDP reaches 0.715 on 30 held-out statements from ten labelled patients each, with 83k parameters trained only on PTB-XL. The label sets differ. This is, therefore, context rather than a ranking, but it places SIDP in the same range as much larger pretrained models.

We hope these arguments address your concerns, and we would be grateful if you would consider them in your assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.