acceptodds
Under review as a conference paper at ICLR 2027

Learning Where to Adapt: Learned Subspace Test-Time Adaptation for Medical Time Series

Abstract

Test-time adaptation (TTA) promises to personalize medical time series classifiers to a new subject from a few unlabeled recordings. Its standard recipe, entropy minimization over normalization parameters, follows the model's uncertainty rather than the ways in which subjects actually differ, and with few windows it can turn correct predictions into errors. We argue that where to adapt should itself be learned. Because inter-subject shift is plausibly driven by a few shared physiological and recording factors, we propose Learned Subspace Test-Time Adaptation (LSTA), which meta-learns from past subjects a low-rank space of LayerNorm updates and restricts each new subject's adaptation to it. During meta-training, unlabeled support windows determine an entropy step within this space, and the classification loss on labeled query windows from the same subject shapes it. At test time the space is fixed, the parameters outside LayerNorm remain frozen, and each subject adapts from at most eight unlabeled windows. On five EEG and ECG benchmarks with subject-exclusive splits, LSTA obtains the highest mean accuracy and macro-F1 among seven TTA baselines, and a learned subspace outperforms a random one in 22 of 25 dataset-seed pairs. Across all eight datasets we evaluate, LSTA is the only method whose macro-F1 never falls below that of the frozen model, and under contaminated support it degrades far less than direct LayerNorm adaptation and turns fewer correct predictions into errors. On clean data, its gain over direct LayerNorm adaptation is concentrated on one dataset.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Test-Time Adaptation, Medical Time Series, Meta-Learning, Low-Rank Subspace, Inter-Subject Shift, EEG, ECG, LayerNorm Adaptation

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.40 over 5 reviews, above 68% of the others.
Reviewer_ADC6OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies test-time adaptation for medical time series from new subjects. The authors argue that standard entropy minimization updates all LayerNorm parameters along directions of model uncertainty and, when only a few windows are available, can turn correct predictions into incorrect ones. They propose LSTA: a rank-8 LayerNorm update subspace BB is meta-learned on training subjects. Support windows are used for one entropy-minimization step within this subspace, and the classification loss on query windows is used to learn BB. At test time, BB is fixed, and each new subject receives one projected-gradient update using at most eight unlabeled windows. With TeCh as the backbone, LSTA achieves higher average macro-F1 than seven TTA methods across five datasets with subject-level labels and is more stable under settings such as contaminated support windows.

Critical strengths

The problem is clearly defined and practically relevant: personalizing a model for a new subject using a few unlabeled windows. The method is simple, requiring only one additional projected-gradient step at inference time, with overhead comparable to Tent.

Critical weaknesses
  1. The gains are small and not statistically significant: LSTA scores 79.23 versus 78.19 for LN-TTA (p≈0.054p \approx 0.054), with improvements concentrated mainly on ADFTD. The simple Frozen-PA baseline outperforms LSTA on PTB but is omitted from Table 1. On Sleep-EDF, the “never below frozen” claim rests on a margin of just 0.06.
  2. The novelty is limited: BB⊤BB^\top is essentially a meta-learned low-rank preconditioner, similar in spirit to Meta-Curvature, CAVIA [1], and LEO [2]. The central premise that subject differences are low-rank is not directly validated.

[1]: Luisa Zintgraf, Kyriacos Shiarlis, Vitaly Kurin, Katja Hofmann, and Shimon Whiteson. Fast Context Adaptation via Meta-Learning. In Proceedings of the 36th International Conference on Machine Learning (ICML), 2019.
[2]: Andrei A. Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. Meta-Learning with Latent Embedding Optimization. In International Conference on Learning Representations (ICLR), 2019.

Questions for authors
  1. Could the authors include Frozen-PA in Table 1? What happens when LSTA is combined with prediction aggregation?

  2. Could the authors test the low-rank assumption directly? For example, they could fine-tune LayerNorm separately for each test subject using labels to obtain oracle updates, then examine the effective rank of those updates and how much of their variance the learned BB explains.

  3. To separate the effect of the learned directions from that of the low-rank constraint, could the authors add a learnable full-rank preconditioner, as in Meta-Curvature, and a CAVIA-style comparison [1] that adapts context parameters?

  4. Are the two pools split randomly or into contiguous time segments? Could the authors add a chronological setting in which the first kk windows are used for adaptation and later windows are predicted? This would more closely reflect deployment.

  5. The experiments use only TeCh as the backbone. Do the conclusions hold with Medformer or a pretrained model such as LaBraM?

  6. Could the authors provide paired tests at the subject level rather than relying only on sign tests over dataset–seed pairs?

[1]: Luisa Zintgraf, Kyriacos Shiarlis, Vitaly Kurin, Katja Hofmann, and Shimon Whiteson. Fast Context Adaptation via Meta-Learning. In Proceedings of the 36th International Conference on Machine Learning (ICML), 2019.

6 comments
AuthorsOpenReview ↗

1/3 Response to Reviewer ADC6

We thank the reviewer for the concrete, testable suggestions and for finding the paper "clear and easy to follow." We ran every experiment in the way you suggested. We will upload the revised paper with changes marked in blue. The method is unchanged in every experiment. Scores are macro-F1; intervals are 95% paired bootstrap over test subjects.

W1 / Q1: Significance, Frozen-PA and Sleep-EDF.

On clean data, LSTA's gain over LN-TTA is significant at the subject level (pooled over seven datasets, p = 0.0017). LSTA is the best method overall, and its advantage is largest where adaptation is unreliable:

LSTA minus …Significant gains (95% subject-level interval)
FrozenAPAVA +7.53 [+1.61, +16.74], TDBrain +6.53 [+1.99, +14.68], ADFTD +6.95 [+1.16, +12.95], PTB-XL +2.27 [+1.85, +2.68]
LN-TTAADFTD +3.89 [+0.81, +6.81], MIT-BIH +1.18 [+0.11, +2.10], Sleep-EDF +0.14 [+0.01, +0.27]; pooled over seven datasets p = 0.0017
  • Best overall on clean data. LSTA has the highest average macro-F1 (79.23 vs. 76.93 for the best baseline) and the gain over LN-TTA holds on the largest cohort, CODE-15% (9,729 patients): +1.17 [+0.83, +1.52].

  • Significant gains under unreliable support. With 50% contaminated support LSTA beats LN-TTA by up to 11.8 points (APAVA 77.41 vs. 65.62), cuts confident errors from 17.6% to 0.1% on PTB-XL, and is significantly more robust on 5 of 6 backbone–dataset pairs.

  • At no extra cost. One gradient step on 8 unlabeled windows, the same FLOPs as Tent (8–21 ms on a GPU).

Frozen-PA. Our main results table compared published adaptation methods, so Frozen-PA which only averages the frozen model's predictions over a subject's windows, was reported in the support-size analysis. We have added Frozen-PA and LSTA + PA to the main results table:

APAVATDBrainPTBADFTDPTB-XL
Frozen-PA88.9794.9484.5451.5564.65
LSTA93.1898.8884.0955.3264.66
LSTA + PA93.3398.9585.3655.2965.22
  • Frozen-PA is a strong baseline, but LSTA is higher overall: 79.23 vs. 76.93 on average, and significantly higher on APAVA (+4.27), TDBrain (+3.94) and ADFTD (+3.83).
  • PTB and the combination. On PTB, LSTA and Frozen-PA are statistically level (p = 0.76), and combining them works best: LSTA + PA reaches 85.36 on PTB and is the best entry on four of the five datasets (mean 79.63).
  • PA assumes one label per recording, so it is not suited where labels change within a recording.
  • Sleep-EDF. We revised the 'never below the frozen model' wording in the paper. Sleep-EDF has little room for adaptation (frozen 72.36), and LSTA is the only adaptation method that stays above the frozen model (72.42), while the other methods lose up to 5.2 points. Its gain over LN-TTA there is also significant at the subject level (+0.14).

W2: Novelty and the low-rank premise.

We agree with the description: LSTA is a meta-learned low-rank preconditioner, and we stated in the revised related work that we do not claim a new adaptation primitive. The difference to Meta-Curvature, CAVIA and LEO is the setting in which the step is taken: they adapt on a labeled support set and mostly meta-train the network itself. LSTA adapts on a few unlabeled windows of a new patient, on an already-trained frozen model, which is the setting of clinical deployment. Our contribution is to show that in this setting the learned low-rank restriction is what makes the unlabeled step safe: learned full-rank (Meta-Curvature) and CAVIA-style rules trained under the same protocol are significantly less robust under contaminated support: by 5.7 to 27.2 points for Meta-Curvature, and by 10.9 to 15.6 points for the CAVIA-style rule on APAVA and TDBrain (Q3).

The premise is now tested directly, as you proposed (Q2): subject-specific updates have an effective rank below 2, and a rank-8 PCA subspace estimated from training subjects contains 93–98% of unseen subjects' updates (random rank-8: below 0.3%).

AuthorsOpenReview ↗

2/3 Response to Reviewer ADC6

Q2: Low-rank assumption.

As you proposed, for each subject we computed the labeled LayerNorm update at the frozen model (class mean removed), and fitted an oracle update with that subject's labels.

APAVATDBrainADFTD
Effective rank of subject updates1.61.21.8
Unseen subjects' variance inside rank-8 PCA of training subjects92.7%97.2%97.8%
… inside a random rank-8 subspace0.06%0.23%0.13%
… inside B16.2%9.1%2.3%
Oracle gain over frozen kept inside B / random rank-8100% / 58%100% / 92%99.6% / 84%
  • The low rank and shared assumption holds. Subject updates have effective rank below 2, and test subjects' updates lie almost entirely in a subspace learned from training subjects.
  • What B explains. Although B holds only 2–16% of variance, that restriction keeps 99.6–100% of the gain, while a random rank-8 subspace keeps 58–92%. B captures the part of the subject difference that changes predictions, not its largest-variance part, because it is trained for safe unlabeled steps rather than reconstruction.
  • A variance basis is less safe. Replacing B with the top-8 PCA basis of the oracle updates is on par on clean data, but significantly less robust under 50% contamination: B is better by +10.75 (APAVA), +2.10 (TDBrain) and +2.07 (PTB-XL).

Q3: Full-rank preconditioner and CAVIA-style comparison

We kept each update operator and trained it under our unlabeled protocol on the same frozen backbone. Mean over 6 datasets:

Clean50% contaminated
LN-TTA69.9159.54
Learned diagonal preconditioner70.3659.86
Random rank-8 subspace69.6659.91
LSTA70.8264.28

We then ran the two under our protocol (unlabeled step, frozen backbone, step size and learning rate chosen on validation): Meta-Curvature with a full-rank matrix starting at identity, and CAVIA with an unconstrained linear map from 8 context parameters into LayerNorm. 50% contaminated support, 3 seeds:

LN-TTALSTAMeta-CurvatureCAVIA-style
APAVA65.1878.0863.3061.69
TDBrain88.2689.4361.8878.51
ADFTD40.9141.5335.9139.26

On clean data all three are within one point except on ADFTD (55.94 vs. 52.92 and 54.03, not significant). This separates the effects you ask about:

  • Low-rank restriction. LSTA is significantly better than Meta-Curvature on all three datasets (+5.7 to +27.2) and than the CAVIA-style rule on two (+10.9 to +15.6). Due to restricting unlabeled step to a learned low-rank subspace.
  • Learned directions. A random subspace is 4.4 points lower under contamination, and lower on clean APAVA (91.07 vs. 93.18), PTB (81.47 vs. 84.09) and ADFTD (51.08 vs. 55.32).
  • CAVIA adapts its context parameters on a labeled support set and meta-trains the whole network. When we replace that with an unlabeled step, as our setting requires, it loses most of its robustness. What LSTA adds is making this kind of low-rank adaptation safe without labels, on a frozen model.
AuthorsOpenReview ↗

3/3 Response to Reviewer ADC6

Q4: Pools and chronological setting.

The pools are the first and second half of each recording in time, so they are contiguous; support comes from one half, the other half is predicted, and the roles are then swapped. We evaluated on the requested setting (LSTA minus frozen; bold: 95% subject-level interval excludes 0 for rows 1, 3 and 4; PTB-XL: 300-subject subset):

SettingAPAVATDBrainPTB-XL
Both halves, both directions (paper protocol)+7.53+6.53+2.20
First half → predict second half+6.81+4.52+1.93
Earliest 8 windows → predict all later+6.96+6.37+3.65
8 windows from first 10% → predict remaining 90%+12.54+5.37+2.78
  • The gain holds when only early windows are used: LSTA is above the frozen model in every chronological setting, by up to +12.54.
  • Later visits. On PTB-XL, adapting on a patient's earliest ECG and predicting that patient's later ECGs (315 patients, up to 4.7 years later) gives +6.86 [+4.82, +9.07] over the frozen model, and +8.20 [+1.22, +13.75] more than a year later. On two external cohorts, with no fitting on either cohort: +3.84 [+2.50, +5.33] (MIMIC-IV-ECG, 1,662 patients) and +3.37 [+2.97, +3.77] (CODE-15%, 9,729 patients).

Q5: Medformer and Transformer

We tested our method on backbones: Medformer and Transformer. Δ\Delta(LSTA - LN-TTA) on 50% contaminated support:

BackboneAPAVATDBrainPTB-XL
Transformer+11.64+3.90+5.93
Medformer+22.40+5.11−0.26
  • Robustness carries over to both backbones: LSTA is significantly more robust than LN-TTA on 5 of 6 backbone–dataset pairs, by up to 22.40 points, and on none of them significantly below the frozen model.
  • Clean gains: LSTA is above the frozen model on all 6 pairs (Medformer: APAVA +7.54, TDBrain +5.59, PTB-XL +1.91; Transformer: APAVA +6.65, PTB-XL +1.53, TDBrain +2.14).

Q6: Subject-level tests

We now use paired bootstrap tests over test subjects (10,000 resamples) comparing LSTA with the frozen model, Frozen-PA, LN-TTA and LN-TTA-Learnable on all seven datasets. Δ\Delta (LSTA - Frozen Model) on clean data:

APAVA (4 subjects)TDBrain (8)ADFTD (19)PTB-XL (3,521)
+7.53 [+1.61, +16.74]+6.53 [+1.99, +14.68]+6.95 [+1.16, +12.95]+2.27 [+1.85, +2.68]
  • We confirmed the gain over the frozen model at the subject level, and pooled over seven datasets the clean gain over LN-TTA is significant (p = 0.0017), with significant per-dataset gains on ADFTD (+3.89), MIT-BIH (+1.18) and Sleep-EDF (+0.14).
  • Because some test sets are small, we also ran 5-fold subject-level cross-validation with retrained backbones (3 seeds), so that every subject is tested once. LSTA minus frozen is +7.66 [+4.19, +11.16] on ADFTD (88 subjects) and +2.85 [+1.01, +4.70] on TDBrain (50 subjects), where LSTA is also above Frozen-PA (+1.21 [+0.17, +2.28]).

We thank the reviewer; we will add analysis in the revised paper for each suggestion. We hope these answers address the reviewer's concerns, and we would be happy to answer any further questions.

Reviewer_PA5LOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the classification bias in test-time adaptation (TTA) by proposing Significant-subspace Alignment (SSA) for continuous regression tasks. Recognizing that entropy minimization cannot be applied to scalar outputs, the authors believe that naive feature alignment causes over-correction in deep regression networks due to the tight, narrow geometry of their latent spaces. The proposed method projects test-time features onto a low-dimensional linear subspace and applies a dimension-weighting heuristic to align only the components deemed relevant to the target variable.

Critical strengths
  1. The manuscript correctly pinpoints a systemic reliance on classification metrics (like entropy) in the TTA literature, making the expansion into regression a necessary, if predictable, next step.

  2. The authors successfully document that standard unsupervised domain adaptation routines perform poorly when directly copied into raw deep regression pipelines, providing a clear justification for why specialized constraints are needed.

Critical weaknesses
  1. The technical contribution is thin. Strip away the TTA terminology, and the method is essentially an application of classical Principal Component Analysis (PCA) and basic linear projections to frozen feature maps. Applying well-established statistical techniques to a newer problem setting does not constitute a fundamental algorithmic breakthrough.

  2. The entire alignment mechanism assumes the pre-trained source network has already cleanly mapped out the relevant feature space. If the source model contains any representation collapse or subtle vulnerabilities, the subspace isolation step collapses as well, rendering the entire test-time optimization loop counterproductive or unstable.

Questions for authors
  1. What is the exact computational overhead introduced by executing subspace detection and matrix inversions dynamically during inference? Provide the precise wall-clock latency per test batch compared to standard forward passes to justify its viability for online systems.

  2. Given that deep features reside on complex, non-linear manifolds, a strictly linear subspace alignment is mathematically suboptimal. Why was a linear projection chosen, and exactly how much performance is sacrificed by failing to model non-linear coordinate shifts?

  3. In scenarios where the target distribution shift changes the true underlying correlation between features and labels, how does your dimension weighting prevent catastrophic negative transfer? Where is the mathematical bound showing this method will not drift under extreme covariate shifts?

  4. If the incoming test data arrives with significant noise or missing sequences, how severely does the subspace detection algorithm degrade? Provide a sensitivity analysis mapping alignment failure against increasing input noise levels.

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

The paper introduces LSTA, a meta-learned subspace method for test-time adaptation of medical time-series classifiers. LSTA learns a low-rank basis of LayerNorm parameter updates from training subjects. During meta-training, unlabeled support windows determine a one-step entropy-based adaptation within the span of the basis, and classification loss on separate labeled query windows trains the basis. At test time the basis is fixed, and each new subject adapts using at most eight unlabeled windows. The evaluation covers eight EEG/ECG datasets. LSTA achieves the highest mean macro-F1 and accuracy among the compared TTA methods on five subject-labeled benchmarks.

Critical strengths

The problem is well motivated and practically relevant.

Test-time adaptation can worsen predictions when only a few unlabeled windows are available, and the paper takes this risk seriously through support-contamination experiments, counts of correct-to-incorrect prediction changes, and sweeps over support size, step count and step size. The reported reductions in harmful prediction changes are encouraging, although, as discussed below, their cause is not yet established.

The method is simple and computationally lightweight.

Restricting LayerNorm updates to a learned low-dimensional subspace adds little inference overhead. LSTA outperforms the random-subspace control in 22 of 25 dataset–seed pairs, providing evidence that learning the basis is useful beyond imposing a random low-rank constraint.

Critical weaknesses

The relationship to prior meta-learning approaches could be clarified.

LSTA’s update can be viewed as applying a learned low-rank orthogonal projector to the adaptation gradient. This connects to prior work on learned gradient transformations, such as Meta-Curvature, and learned adaptation subspaces, such as MT-Net (Lee & Choi, ICML 2018). The paper discusses Meta-Curvature but does not compare with it experimentally, and MT-Net is not discussed. The learned parameter-wise scaling baseline is a useful control, but further clarification of these connections would help identify what the proposed formulation adds. The avoidance of second-order derivatives is a useful consequence of evaluating the single-step support gradient at fixed source parameters, independently of the basis. I view the specific formulation for unlabeled subject adaptation as reasonable, but encourage clearer positioning relative to these conceptual predecessors.

The robustness results do not fully separate update direction from update magnitude.

At a fixed learning rate, projecting the entropy gradient onto an orthonormal rank-8 subspace cannot increase its norm. Smaller effective updates could contribute to fewer correct-to-incorrect prediction changes, better preservation of the source representation, and less degradation under contaminated support. Although LN-TTA’s learning rate is selected on validation data, the paper does not report the resulting update norms or compare against LN-TTA with updates rescaled to match LSTA’s norm. Without that control, it remains unclear how much of the robustness benefit comes from learning useful directions beyond making more conservative updates. The observation that Random-Subspace also stays close to the frozen model under contamination on HPP is consistent with this concern, although it does not establish the explanation.

The evidence for clean-data gains over the direct adaptation controls is limited.

LSTA outperforms LN-TTA in 17 of 25 dataset–seed pairs (p ≈ 0.054) and LN-TTA-Learnable in 16 of 25 (p ≈ 0.11), with the largest gain concentrated on ADFTD. These pooled sign tests do not reach the conventional 0.05 significance threshold, although this does not establish that the methods are equivalent. Several test sets are also very small (four test subjects on APAVA and eight on TDBrain) so aggregate differences could be sensitive to a small number of subjects. The reported standard deviations quantify variability across backbone seeds rather than uncertainty across held-out subjects.

The evaluation scope is limited.

LSTA is evaluated on a single backbone family, TeCh. Table 1 also omits frozen prediction aggregation (Frozen-PA), which has slightly higher reported macro-F1 than LSTA on PTB at k = 8 (84.54 vs. 84.09); including it would make the comparison with this simple alternative more transparent. Although training and test subjects are disjoint, adaptation and evaluation use separate windows from the same test subject under an offline, within-recording protocol. This is appropriate for the stated personalization setting, but does not establish whether the adaptation remains useful on later recordings or visits. In addition, maintaining or improving dataset-level macro-F1 does not imply that every subject benefits. While the paper acknowledges several of these limitations, the results should be interpreted as reduced harmful adaptation under the evaluated conditions, rather than a general “do no harm” guarantee.

Questions for authors
  1. Can you report the distribution of LayerNorm update norms for each method? Can you also compare LSTA against LN-TTA with its updates rescaled to match LSTA’s norm on the same support episodes, on clean data, under contaminated support, and for harmful prediction changes? This is the experiment most likely to change my assessment.

  2. Can you report the fraction of test subjects whose performance improves or deteriorates after adaptation, together with the distribution of these changes using an appropriate subject-level metric? This is especially relevant for APAVA and TDBrain, given their small test sets.

  3. Could you include Frozen-PA in Table 1? Could you also clarify how LSTA relates to MT-Net and Meta-Curvature, and whether either could provide a meaningful comparison under your protocol?

  4. Do you have any evidence that the results hold with backbones other than TeCh?

5 comments
AuthorsOpenReview ↗

1/2 Response to Reviewer pvTq

We thank the reviewer for the careful and precise review. We appreciate that you found the problem "well motivated and practically relevant," that the paper "takes this risk seriously" through robustness experiments, analyses and parameter sweeps. We ran the experiment you proposed to find the cause of reduced harm, "the experiment most likely to change my assessment," exactly as described. It showed that at equal update size, LSTA is significantly more robust than LN-TTA under contaminated support and makes fewer harmful changes, so the benefit comes from the learned direction, not from a smaller step. We will upload the revised paper with all changes marked in blue. Scores are macro-F1, and bold marks differences whose 95% paired bootstrap interval over test subjects excludes 0.

W2 / Q1: Direction versus magnitude.

(i) Update norms. Median relative update size ∥ϕ−ϕ0∥/∥ϕ0∥\|\phi - \phi_0\| / \|\phi_0\| per test episode, each method at its own validation-selected step size:

APAVATDBrain
LSTA0.0290.033
LN-TTA0.1450.011
Random-Subspace2.2302.456

Because each method selects its own step size, the projection does not make LSTA's update smaller: on TDBrain it is 3× larger than LN-TTA's.

(ii) LN-TTA rescaled, per episode, to exactly LSTA's norm on the same support (LSTA minus rescaled LN-TTA; bold: LSTA significantly better, 95% subject-level interval):

Clean25% contaminated50% contaminated
APAVA+0.11+0.29+3.71
TDBrain−0.04+0.90+12.09
PTB-XL−0.31−0.21+1.44
  • Under contamination, the learned direction is what makes LSTA robust: at equal size it is significantly better on all three datasets at 50%, and on APAVA and TDBrain already at 25%.
  • On clean data, the two directions perform alike at equal size: within 0.31 points, with rescaled LN-TTA slightly ahead on TDBrain (−0.04) and PTB-XL (−0.31).

(iii) Harmful / useful prediction changes at equal norm (% of frozen-correct windows turned wrong / % of frozen-wrong windows turned right):

LSTALN-TTA@LSTARandom@LSTA
APAVA, clean2.53 / 69.72.77 / 70.50.27 / 15.4
APAVA, 50% contam.13.45 / 31.418.04 / 33.70.95 / 4.5
TDBrain, clean0.44 / 91.30.66 / 94.40.10 / 46.4
TDBrain, 50% contam.5.04 / 36.019.05 / 46.80.86 / 9.6
PTB-XL, clean3.32 / 17.05.34 / 22.21.73 / 8.7
PTB-XL, 50% contam.3.61 / 11.88.32 / 17.91.46 / 5.4
  • At the same size, LSTA turns fewer correct predictions into errors in all six settings, by up to 4× under contamination (TDBrain 5.04% vs. 19.05%).
  • Size alone does not explain LSTA's behaviour: random directions at exactly LSTA's size barely change predictions (APAVA: 15.4% of errors corrected vs. 69.7%).

In short: under contaminated support, the robustness comes from the learned direction, not from a smaller step; on clean data, LSTA and LN-TTA are equivalent at equal update size. The learned subspace therefore costs nothing significant on clean data and protects when the support is unreliable.

W1 / Q3: MT-Net and Meta-Curvature. We agree with your reading: LSTA applies a learned low-rank orthogonal projector BB⊤BB^\top to the entropy gradient, and it stays first-order because that gradient is taken at the fixed source parameters. Revised related work now discusses both methods. Both take their inner step on a labeled support set and meta-train the network; a new patient has no labels, so we kept each update operator and trained it under our protocol (unlabeled entropy step, same frozen backbone, step size and learning rate selected on validation).

  • Meta-Curvature (a learnable full-rank matrix per LayerNorm tensor, starting at identity) is within one point of LSTA on clean data except ADFTD, but under 50% contaminated support LSTA is significantly better: APAVA 78.08 vs. 63.30 (+14.08 [+5.26, +23.17]), TDBrain 89.43 vs. 61.88 (+27.21 [+14.64, +43.54]) and ADFTD 41.53 vs. 35.91 (+5.67 [+1.79, +10.86]).
  • An MT-Net-style update (learned T-matrices and mask) is lower than LSTA both on clean data and under contamination.
  • What the formulation adds: a learned full-rank transformation fits clean data as well as LSTA but is not robust without labels. Restricting the unlabeled step to a learned low-rank subspace is what makes it safe.
AuthorsOpenReview ↗

2/2 Response to Reviewer pvTq

W3 / Q2: Subject-level statistics.

Revised paper will replaces the sign tests with paired bootstrap tests over test subjects (10,000 resamples), comparing LSTA with the frozen model, Frozen-PA, LN-TTA and LN-TTA-Learnable on all seven datasets.

(i) Gains over the frozen model Δ\Delta (LSTA - Frozen):

APAVATDBrainADFTDPTB-XL
+7.53 [+1.61, +16.74]+6.53 [+1.99, +14.68]+6.95 [+1.16, +12.95]+2.27 [+1.85, +2.68]
  • The clean-data gain over direct adaptation is significant. Pooled over the seven datasets, LSTA beats LN-TTA (p = 0.0017) and LN-TTA-Learnable (p = 0.028), with significant per-dataset gains on ADFTD (+3.89 [+0.81, +6.81]), MIT-BIH (+1.18 [+0.11, +2.10]) and Sleep-EDF (+0.14 [+0.01, +0.27]) with LN-TTA.

(ii) Small test sets. We kept the published subject-exclusive splits of TeCh and Medformer for comparability, and added 5-fold subject-level cross-validation with retrained backbones:

  • The gains hold on larger test sets: ADFTD +7.66 [+4.19, +11.16] (88 subjects) and TDBrain +2.85 [+1.01, +4.70] (50 subjects).

(iii) Per-subject changes vs. frozen (improved / unchanged / worsened):

Dataset (subjects)LSTALN-TTA
ADFTD, cross-validation (88)71.6 / 0.0 / 28.4%64.8 / 1.1 / 34.1%
MIT-BIH (22)50.0 / 40.9 / 9.1%45.5 / 31.8 / 22.7%
  • LSTA improves more subjects and worsens fewer than LN-TTA on ADFTD and MIT-BIH.
  • On APAVA and TDBrain, which you asked about, no test subject gets worse with LSTA (APAVA 75.0 / 25.0 / 0.0%; TDBrain 50.0 / 50.0 / 0.0%).

We agree with your reading of "do no harm": The revised paper will now state the measured result, reduced harmful adaptation under the evaluated conditions.

W4a: Frozen-PA.

As Frozen-PA is simple aggregation rather than a published method it was reported in support size analysis. We added to the main results table together with LSTA + PA (bold: highest).

APAVATDBrainPTBADFTDPTB-XL
Frozen-PA88.9794.9484.5451.5564.65
LSTA93.1898.8884.0955.3264.66
LSTA + PA93.3398.9585.3655.2965.22

LSTA is better than Frozen-PA on APAVA (+4.27), TDBrain (+3.94) and ADFTD (+3.83), and level on PTB, where LSTA + PA (85.36) is above Frozen-PA (84.54).

W4b: Later recordings and visits

  • Later visits. On PTB-XL, adapting on a patient's earliest ECG and predicting that patient's ECGs from later dates (315 patients; median 26 days, up to 4.7 years later) improves macro-F1 by +6.86 [+4.82, +9.07] over the frozen model, with less than half of LN-TTA's harmful changes (2.7% vs. 6.0%). The gain holds for ECGs recorded more than a year later (+8.20 [+1.22, +13.75]). With 50% contaminated support LSTA is +12.86 [+11.00, +14.79] ahead of LN-TTA, and 12–16 points ahead in time range.
  • Two external cohorts, same backbone. MIMIC-IV-ECG (1,662 patients with an unchanged diagnosis, median 388 days later): +3.84 [+2.50, +5.33] over the frozen model, with 5.0% harmful changes against 10.8% for LN-TTA. CODE-15% (all 9,729 patients with two or more labeled exams): +3.37 [+2.97, +3.77] over the frozen model and +1.17 [+0.83, +1.52] over LN-TTA; with 50% contaminated support, +17.45 [+17.03, +17.88] over LN-TTA (harmful changes 2.7% vs. 23.5%).
  • Later in the same recording. Adaptation and evaluation already use different halves of each recording in time. With 8 windows from the first 10% of the recording, predicting the remaining 90%, LSTA improves over the frozen model on APAVA (+12.54), TDBrain (+5.37) and PTB-XL (+2.78).

Q4: Other backbones . We applied LSTA unchanged to a vanilla Transformer and Medformer, both trained with the Medformer authors' code. LSTA minus LN-TTA under 50% contaminated support:

BackboneAPAVATDBrainPTB-XL
Transformer+11.64+3.90+5.93
Medformer+22.40+5.11−0.26
  • Robustness carries over to both backbones: LSTA is significantly more robust than LN-TTA on 5 of 6 backbone–dataset pairs, by up to 22.40 points, and on none of them significantly below the frozen model.
  • Clean gains: LSTA is above the frozen model on all 6 pairs (Medformer: APAVA +7.54, TDBrain +5.59, PTB-XL +1.91; Transformer: APAVA +6.65, PTB-XL +1.53, TDBrain +2.14).

We thank the reviewer again; the requested control substantially sharpened what the paper claims. We hope these answers address the reviewer's concerns, and we would be happy to answer any further questions.

AuthorsOpenReview ↗

Followup on Rebuttal

Dear Reviewer,

Thank you for your thoughtful feedback, particularly your questions about the source of LSTA’s robustness. As the rebuttal period draws to a close, we would greatly appreciate it if you could revisit our responses and let us know whether any concerns remain unresolved.

In the above response, we have added the requested norm-matched controls separating update direction from magnitude, subject-level analyses and cross-validation, comparisons with Frozen-PA and prior meta-learning approaches, and evaluations on additional backbones and later recordings, including external cohorts. We have also clarified the scope of our claims regarding reduced harmful adaptation.

We hope these answers address the reviewer's concerns, and we would be happy to answer any further questions.

Thank you again for your time.

Reviewer_fC25OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Learned Subspace Test-Time Adaptation (LSTA) for medical time-series classification under inter-subject distribution shifts. The method meta-learns a shared low-rank subspace of LayerNorm parameter updates from training subjects. At test time, the model adapts to unseen subjects using entropy minimization on a small number of unlabeled windows, with updates restricted to the learned subspace.

The authors evaluate LSTA on eight EEG and ECG datasets and report improvements over several test-time adaptation baselines, particularly in robustness to contaminated support samples.

Critical strengths
  1. The paper addresses a practically relevant problem in medical time-series classification, where distribution shifts between subjects can degrade performance and labeled data from new subjects are often unavailable.

  2. The experimental evaluation covers multiple EEG and ECG datasets and includes comparisons with several test-time adaptation methods. The additional robustness analyses, particularly those involving contaminated support samples and harmful prediction changes, provide useful empirical observations.

  3. The proposed method is relatively straightforward to implement and introduces limited additional inference overhead.

Critical weaknesses

1. The central assumption of a shared low-dimensional space for inter-subject variation is not substantiated.

The main motivation of this paper is that inter-subject distribution shifts in medical time series can be characterized by a shared low-dimensional space of parameter changes. I consider this an important and potentially interesting hypothesis. If validated, it could provide meaningful insights into the structure of inter-subject variability in medical data.

However, the paper does not provide direct evidence supporting this assumption. The experiments demonstrate that restricting parameter updates to a learned low-rank subspace can improve classification performance, but this does not establish that inter-subject variations actually exhibit such a low-dimensional structure.

More importantly, the adaptation basis is learned by minimizing the supervised classification loss on query samples. Therefore, the learned subspace may simply function as a task-optimized gradient projection that suppresses harmful updates, rather than capturing the underlying structure of inter-subject variation.

For example, if the learned space genuinely represents shared inter-subject variation, one might expect subjects exhibiting similar distribution shifts to require similar adaptation directions or coefficients. However, the paper does not investigate this relationship or directly analyze the dimensionality of subject-specific parameter changes.

Consequently, the experimental evidence supports the effectiveness of the proposed optimization constraint, but not the central scientific hypothesis used to motivate the method. I believe this is a substantial gap between the paper's motivation and its actual findings.

2. The methodological novelty is limited, and the marginal improvement over a simple baseline makes the practical contribution difficult to justify.

The technical approach is conceptually similar to existing feature-wise modulation methods such as FiLM, which adjust intermediate feature representations through learned scaling and shifting parameters. LSTA similarly modifies the affine parameters of LayerNorm layers, but uses meta-learning to learn a low-rank space of allowable updates and entropy minimization to determine subject-specific adjustments.

Although the optimization procedure differs from FiLM, the underlying idea of adapting feature representations through affine parameter modulation is not new. Furthermore, learning low-dimensional parameter update spaces and gradient transformations has already been explored in meta-learning literature. The paper does not sufficiently clarify what fundamentally new adaptation mechanism is introduced beyond combining these established techniques.

More importantly, the experimental results raise questions about the practical significance of this additional mechanism. According to Table 2, direct LayerNorm test-time adaptation (LN-TTA) already improves the average macro-F1 from 74.10% to 78.19%. LSTA further increases this result to 79.23%, representing an improvement of only 1.04 percentage points over LN-TTA.

The improvement is primarily concentrated on ADFTD, while the differences on the remaining datasets are relatively small or even slightly negative. The reported comparison between LSTA and LN-TTA is also not statistically significant at the conventional 0.05 level (17 out of 25 dataset-seed pairs, p ≈ 0.054).

Given that a simple entropy-based LayerNorm adaptation strategy already achieves most of the performance gain, it is difficult to determine whether learning the proposed subspace provides a sufficiently substantial and consistent improvement to justify its additional meta-training procedure.

In its current form, the method appears closer to an incremental refinement of existing feature modulation and test-time adaptation techniques than a fundamentally new approach. Without stronger evidence of a distinct underlying mechanism or more substantial improvements over simple baselines, I find it difficult to assess the broader value of this contribution to the community.

Overall Assessment

The paper addresses a relevant problem and presents reasonably comprehensive experiments. The robustness results are valuable, and the proposed approach is technically straightforward.

However, two issues substantially affect my assessment. First, the central hypothesis that inter-subject variations share an intrinsic low-dimensional structure is not directly validated. Second, the proposed technical approach is closely related to established feature modulation and subspace-based adaptation methods, while the improvement over direct LayerNorm adaptation is only 1.04 percentage points on average and is not consistently observed across datasets.

As a result, I find the scientific justification and methodological contribution insufficiently compelling in the current submission. Stronger evidence supporting the central hypothesis and a clearer demonstration of the necessity and advantages of the learned subspace would be needed to change my assessment.

Questions for authors
  1. Can the authors provide direct evidence that inter-subject variations in medical time series lie in a shared low-dimensional parameter space? For instance, have they analyzed the intrinsic dimensionality of subject-specific parameter updates or examined whether subjects with similar distribution shifts exhibit similar adaptation directions?

  2. How does LSTA fundamentally differ from existing feature-wise affine modulation and subspace-based meta-learning approaches? Given that LN-TTA already achieves most of the performance gain, what evidence demonstrates that learning the proposed subspace is necessary rather than simply beneficial on a limited number of datasets?

  3. The figures are difficult to follow and do not clearly communicate the proposed methodology. The visual presentation of the paper needs improvement, particularly Figure 2, which is intended to illustrate the main framework. Despite separating meta-training and test-time adaptation into two panels, the figure is difficult to interpret due to its complicated layout, dense connections, and insufficiently explained intermediate operations. In particular, it is not immediately clear how the entropy gradients from the support set are projected through the learned basis, how the query loss updates the basis during meta-training, or which parameters remain frozen at different stages. These are essential steps for understanding the proposed method, but the current visualization makes the reader rely heavily on the mathematical descriptions to reconstruct the workflow. I suggest redesigning Figure 2 with a clearer sequential structure, explicitly distinguishing the support-set adaptation step from the query-set meta-optimization step, and highlighting the trainable and frozen components. A more intuitive visualization would substantially improve the readability of the paper.

7 comments
AuthorsOpenReview ↗

1/2 Response to Reviewer fC25

We thank the reviewer for finding the problem "practically relevant" and the evaluation, which covers "multiple EEG and ECG datasets" with robustness analyses to "provide useful empirical observations," and the "limited additional inference overhead". We are also grateful that you called our central hypothesis "important and potentially interesting". We agree that the main motivation needed more evidence, and your suggestions gave us a clear path for testing them. Our analysis supports the hypothesis. We will upload the revised paper with all changes marked in blue. Scores are macro-F1; brackets give 95% paired bootstrap intervals over test subjects.

W1 / Q1: Direct evidence for a shared low-dimensional structure.

For every subject, we computed the gradient of that subject's labeled loss with respect to the LayerNorm parameters, and removed the class mean to obtain subject-specific updates:

DatasetEffective directionsUnseen subjects' updates inside 8 directions from training subjects… inside 8 random directions
APAVA1.692.7%0.06%
TDBrain1.297.2%0.23%
ADFTD1.897.8%0.13%

The results support the hypothesis on both questions you raised:

  • Low-dimensional. Subject-specific updates use only 1.2–1.8 effective directions (participation ratio) out of 2,560–12,288 LayerNorm parameters.
  • Shared across subjects. Eight directions estimated on training subjects contain 92.7–97.8% of unseen subjects' updates, against 0.06–0.23% for eight random directions.
  • Adaptation directions are subject-specific and track the shift. As you suggested, we examined adaptation directions directly. A subject's support-window gradients, projected onto 8 directions, agree across its own episodes (cosine 0.53–0.95) and differ from other subjects' (−0.21 to 0.48). Under synthetic shifts of increasing severity, the difference in adaptation coefficients grows with severity (rank correlation 0.35–0.60).

W1 (cont.): What the learned basis represents.

Low-rank, shared structure is a property of subject-specific updates (above), and B is a set of learned prediction-relevant directions. A labeled update restricted to B recovers 99.6%-100% of the oracle's macro-F1 gain on APAVA, TDBrain and ADFTD.

The measured structure is directly usable, and B adds robustness. We replaced B with the top-8 PCA directions of training subjects' fitted labeled updates (no task training), with the step size selected on validation. LSTA minus the PCA basis (bold: significant):

APAVATDBrainADFTD
Clean+0.09+0.03+3.05
50% contaminated+10.75+2.10+0.84

On clean data the PCA basis is close to B on APAVA and TDBrain (B is better on ADFTD, +3.05), so the measured structure already supports adaptation on its own; under contaminated support B is significantly more robust on APAVA and TDBrain, by up to 10.75 points. The structure explains why 8 directions suffice; the query loss chooses the 8 that are safest to use without labels.

AuthorsOpenReview ↗

2/2 Response to Reviewer fC25

W2a / Q2: Novelty.

We agree that LSTA's parameterization is closer to FiLM, and we explicitly mentioned it in the revised related work . Our contribution is not the parameterization but the problem it solves and how. FiLM adapts from a known conditioning input, and meta-learned update spaces (Meta-SGD, MT-Net, Meta-Curvature) adapt from a labeled support set. In our settings new patient provides neither: only 8 unlabeled windows, some of which may be unreliable. LSTA addresses this by learning where an unlabeled update is safe: it meta-learns, from past subjects, a low-rank subspace that restricts the entropy step on an already trained, frozen model, so that this step improves accuracy on held-out labeled windows. Our experiments show that this restriction, not the parameterization, is what makes unlabeled adaptation robust: existing update rules lose their robustness under the same protocol, while LSTA keeps it. To test this, we kept each method's update operator and trained it in similar setting:

Update rule (same frozen backbone, same unlabeled step)Clean50% contaminated
Direct LayerNorm adaptation (LN-TTA)69.9159.54
Learned per-parameter step sizes (Meta-SGD-like)70.3659.86
Learned modulation network (FiLM-like)70.7658.28
MT-Net-style update61.8653.26
Random rank-8 subspace69.6659.91
LSTA (learned rank-8 subspace)70.8264.28
  • On clean data, LSTA is the highest, with the other learned rules within about one point, except the MT-Net-style update, which is lower.
  • Under contamination, only the learned low-rank subspace stays robust. It is 4.4 to 11.0 points above the other learned rules and 4.4 points above a random rank-8 subspace. Learned step sizes stay within 0.4 points of plain LN-TTA, and the FiLM-like network falls below it.
  • Meta-Curvature, a learned full-rank transformation trained the same way, is 5.7 to 27.2 points below LSTA under contamination on APAVA, TDBrain and ADFTD (e.g. TDBrain 61.88 vs. 89.43).

So the new element is not the parameterization but where the adaptation is allowed to go: a subspace learned from past subjects that makes an unlabeled step safe.

W2b / Q2: Is the learned subspace necessary?

Yes. The learned subspace gives a significant gain on clean data and a large, consistent gain when the support is unreliable.

  • Clean data. Tested over subjects rather than seeds, LSTA's gain over LN-TTA is significant (pooled over seven datasets p = 0.0017), and holds on CODE-15% (9,729 patients): +1.17 [+0.83, +1.52].
  • Contaminated support. With half of the support windows contaminated, LSTA is clearly above LN-TTA on every dataset we tested with our main backbone.
50% contaminated supportAPAVATDBrainADFTD
LSTA − LN-TTA+11.79+1.39+4.01
  • Atypical support. With the patient's own 8 most uncertain windows as support, LSTA never falls more than 0.05 points below the frozen model, while LN-TTA falls 9.1 points below it on APAVA.
  • The meta-training cost is small. It is done once per model, offline: about 7–39 minutes in total, and adapting a new patient then takes one forward and backward pass (8–21 ms on a GPU).

Since support windows cannot be checked before adapting in clinical use, robustness to unreliable support is where the learned subspace is necessary.

Q3: Figure 2. We agree, and we have redesigned Figure 2 for the revision as you suggested: with a clearer sequential structure, explicitly distinguishing the support-set adaptation step from the query-set meta-optimization step, and highlighting the trainable and frozen components. The new figure shows how the entropy gradients from the support set are projected through the learned basis, how the query loss updates the basis during meta-training, and which parameters remain frozen at each stage.

Your two conditions. You wrote that "stronger evidence supporting the central hypothesis and a clearer demonstration of the necessity and advantages of the learned subspace" would change your assessment.

  • Central hypothesis: subject-specific updates are low-rank (1.2–1.8 effective directions) and shared with unseen subjects (92.7–97.8%, against 0.06–0.23% for random directions).
  • Necessity of the learned subspace: with contaminated support, established learned update rules, including a FiLM-like modulation network, are as fragile as direct adaptation (53.3–59.9 mean macro-F1), while LSTA reaches 64.3. LSTA is also above direct LayerNorm adaptation on APAVA (+11.79), TDBrain (+1.39) and ADFTD (+4.01) .

We thank the reviewer again; the dimensionality analysis and the update-rule comparison made the paper's central claim considerably more precise. We hope these answers address the reviewer's concerns, and we would be happy to answer any further questions.

AuthorsOpenReview ↗

Followup on Rebuttal

Dear Reviewer fC25,

Thank you for your thoughtful and detailed review, particularly your concerns regarding the central low-rank hypothesis and the necessity of the learned subspace. As the discussion period draws to a close, we would greatly appreciate it if you could revisit our responses and let us know whether any concerns remain unresolved.

In our responses, we have provided the analyses you suggested: a direct test of the hypothesis through the effective dimensionality of subject-specific LayerNorm updates and how much of unseen subjects' updates lie in directions estimated from training subjects, and an analysis of whether adaptation directions are subject-specific and track the distribution shift. We also compare the learned basis with a PCA basis of subject updates, and compare LSTA with established learned update rules, including a FiLM-like modulation network, Meta-SGD-like, MT-Net-style and Meta-Curvature updates, all trained under the same unlabeled protocol. We further report subject-level paired tests and robustness under contaminated and atypical support, clarify the novelty relative to feature-wise modulation and meta-learned update spaces. All of these analyses will be included in the revised paper.

We hope our responses and the additional evidence have adequately addressed your concerns. We would be happy to answer any further questions.

Thank you again for your time.

Reviewer_3rHWOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work introduces LSTA, a general method for test-time adaptation in subject-independent transfer settings for medical time series. The authors’ central contribution is a tuning mechanism that focuses on the LayerNorm while involving a meta-training phase that is claimed to transfer well to unseen subjects. The proposed method performs well on benchmarks and shows impressive transfer abilities that also work in corrupted sequence settings, an important robustness result. This work overall is well done, but the communication of the paper is lacking and it currently lacks sufficient motivation from the authors on why this problem setting is important for the medical community.

Critical strengths
  1. The work shows impressive performance relative to baselines, beating out several recently-published works in the space.
  2. The authors present several useful ablations to demonstrate the benefits of their method. The robustness experiments are also appreciated and show performance beyond simple benchmark numbers.
  3. Thank you to the authors for sharing code and experimental details in the appendix. This code seems to contain everything needed to reproduce the implementation as well as re-run certain experiments.
Critical weaknesses
  1. The support selection seems to be problematic from my reading. As authors describe it, they select support sets for tuning randomly throughout the sequence, but this might not transfer to settings such as when the patient’s first part of the sequence is observed without having the last sequence. A natural follow-up would be to select support sets from early in the time series and see if these transfer to later observations that are disjoint temporally rather than randomly chosen.
  2. “Entropy minimization may increase confidence in the wrong class” - this seems to be a very speculative point that isn’t addressed empirically. Further, authors go about optimizing for entropy minimization in Equation 4, so this seems a bit out of place.
  3. The overall motivation of the work is fairly poor. It would be useful to readers to motivate the particular scenarios in which this setting is common in practice. Especially since this work is focused on applications to medical time series, it’s necessary to motivate the settings in the clinic in which this type of method would be used. Without proper motivation, this work seems very niche and incremental.
Questions for authors
  1. It’s not clear that this method is motivated specifically for medical time series, so
  2. Why did authors decide to focus on tuning LayerNorm weights? This is not established or motivated. Is LayerNorm the best set of parameters to update for this problem? If efficiency is the goal, this should be stated explicitly, and comparisons against LoRA would be appreciated.
  3. What is the efficiency of this method? Building off of weakness 3 above, is it practical to expect a realistic setting to allow for this level of tuning. The authors test inference efficiency, but what about adaptation efficiency, e.g., to train the method to adapt to a new subject? Did authors consider how this would be deployed and how the advancements here are motivated in the real-world?
  4. The contaminated support window analysis is interesting. Why do authors think that PTB-XL shows the same performance roughly for LSTA vs. the random-subspace?
7 comments
AuthorsOpenReview ↗

1/2 Reply to Reviewer 3rHW

We thank the reviewer for their thoughtful and constructive feedback, for recognizing the "impressive performance relative to baselines," the "useful ablations," and the robustness experiments that "show performance beyond simple benchmark numbers." and for checking our code." We will upload the revised paper with all changes marked in blue. Scores are macro-F1 over 5 seeds; bold: the 95% subject-level bootstrap interval excludes 0.

W1: Support selection and early windows.

We states it clearly in the revision. Each recording is split in time into a first and a second half (boundary window excluded). Up to 8 support windows are drawn from one half, the other half is predicted, and the roles are then swapped. Because the reversed direction adapts on later windows, we ran the setting you propose: support only from the start of the recording, prediction only on later windows. Δ\Delta (LSTA - frozen model).

Support → predictedAPAVATDBrainPTB-XL
Main protocol (both directions)+7.53+6.53+2.20
First half → second half+6.81+4.52+1.93
Earliest 8 windows → all later+6.96+6.37+3.65
8 from the first 10% → remaining 90%+12.54+5.37+2.78

PTB-XL: 300-subject subset

  • Early support transfers to later windows. The gain stays positive in chronological setting, significantly on APAVA and TDBrain.
  • It persists along the recording. On APAVA the gain holds along the whole recording (+5.2 to +8.4 accuracy points in fifth of it).
  • It transfers to later visits. Adapting on a patient's earliest PTB-XL ECG and predicting that patient's ECGs from later dates (315 patients; median 26 days) improves macro-F1 by +6.86 [+4.82, +9.07], and by +8.20 [+1.22, +13.75] for ECGs recorded more than a year later.

W2: Entropy and wrong-class confidence.

(i) We collected evidences. We restrict it to entropy minimization alone (an unconstrained update). We measure it by share (%) of windows the frozen model classifies correctly that adaptation turns into wrong predictions with confidence above 0.9 (lower is better; bold: lower).

APAVATDBrainPTB-XL
Clean support, LN-TTA (entropy alone)3.590.234.29
Clean support, LSTA2.210.180.35
50% contaminated, LN-TTA (entropy alone)23.384.1917.56
50% contaminated, LSTA10.982.720.08

Entropy minimization alone turns up to 23% of correct windows into confident errors; LSTA nearly removes them (PTB-XL: 17.56% → 0.08%).

(ii) Why Eq. 4 still minimizes entropy. The concern is the unconstrained direction of entropy objective, not the objective itself. Without labels, entropy is the standard signal (Tent, SAR); LSTA applies only the part of its gradient that lies in a subspace learned from past subjects, which filters out most confident errors (table above).

W3/Q1: Clinical motivation.

We revised the introduction to open with a clinical setting. A model is trained once on past patients and then applied to new patients whose signals differ in different aspects(e.g., in gain, baseline and electrical axis). For a new recording, model must predict the diagnosis mitigating shifts introduced by other factors. So no labels are available; what is available is the recording itself, which spans many windows (a resting 12-lead ECG lasts 10 seconds, a clinical EEG several minutes). This gives three requirements:

  1. no labels from the new patient;
  2. no access to the training recordings, which stay at the site that trained the model;
  3. adaptation that rarely turns correct predictions into errors, because the windows are not checked for artifacts before they are used.

Two settings where this is routine:

  • Reading a recording. For an admission ECG or a clinic EEG, the first windows adapt the model and the rest are classified (e.g., 8 one-second EEG windows before hundreds more are read).
  • Repeat recordings of the same patient. Patients are recorded again at follow-up. On PTB-XL, adapting on a patient's earliest ECG improves the prediction of that patient's later ECGs by +6.86 macro-F1 [+4.82, +9.07] (315 patients).

Existing test-time adaptation meets requirements 1 and 2; LSTA mainly targets the requirement 3. With the patient's own 8 most uncertain windows as support, LSTA keeps the frozen model's macro-F1 (LSTA stays above the frozen model on APAVA, TDBrain and PTB-XL (+0.40, +2.23, +1.22), while LN-TTA falls 9.1 below on APAVA), while direct LayerNorm adaptation falls 9.1 points below it on APAVA.

On Q1, the setting is specific to medical time series in three ways: the subject is the natural unit of shift, many past subjects are available to learn the adaptation space from, and each new subject provides only a few unlabeled windows. LSTA is designed around these three properties.

AuthorsOpenReview ↗

2/2 Reply to Reviewer 3rHW

Q2: Why LayerNorm, and LoRA.

Since our main goal is safe adaptation, efficiency is a secondary benefit. We added the following reasons for choosing LayerNorm in the revision::

  1. Normalization parameters are the standard target of test-time adaptation (Tent, ICLR 2021; SAR, ICLR 2023).
  2. They are few (2,560–12,288 in our backbone, 12.5× fewer than rank-8 LoRA), and LSTA further restricts each patient's update to 8 coefficients, so an update estimated from 8 unlabeled windows has little room to overfit.
  3. Among the parameter sets we tested under the same rank-8 formulation, LayerNorm gives the best balance of gain and robustness.

LoRA: LSTA unchanged, with rank-8 LoRA adapters on the encoder's linear layers in place of LayerNorm:

Dataset50% contaminated, LayerNorm50% contaminated, LoRA
APAVA77.4167.64
TDBrain90.4390.02

With contaminated support, LayerNorm is significantly more robust on APAVA in around 10 points and not significantly different on TDBrain; on clean data the two do not differ significantly.

Q3: Adaptation efficiency and deployment.

(i) Adapting to a new patient involves no training. Adaptation takes one backpropagation pass on at most 8 windows, and one projection: 8–21 ms per patient on a GPU (10–190 ms on a CPU). In FLOPs this is same as one Tent step; SAR costs 2×, UCon 4–16× and SHOT about 26× as much. When frozen backbone is available the only training is meta-training, done once per model, offline around 1.7–9.7 minutes per candidate step size, about 7–39 minutes for the four candidates.

(ii) Deployment. The model and the basis needed to be deployed once. When a new recording starts, its first 8 windows give one projected step, and the adapted model classifies the rest of the recording. Adaptation uses only the patient's own unlabeled windows, and the per-patient state is 8 numbers, so a hospital can adapt per patient or per session without storing a model copy per patient and without the training data. An earlier recording of the same patient can also be used.

Q4: PTB-XL, LSTA vs. Random-Subspace.

PTB-XL gives the least subject-specific signal to adapt to. Each 10-second ECG is cut into 1-second windows, so each half of a recording offers only around 4-5 support windows, and with 50% contamination about half of them are foreign. With so little signal, the best an unlabeled method can do is avoid harm: LSTA and the random subspace both stay at the frozen level, while LN-TTA drops far below it.

The equal scores does not show difference between the two bases:

  • LSTA's basis contains the useful directions. A labeled update restricted to it recovers 99.6–100% of the oracle's macro-F1 gain on APAVA, TDBrain and ADFTD, against 58–92% for random directions; a few unlabeled windows can exploit them only partly.
  • The random subspace stays near frozen because its directions barely change predictions. At LSTA's update size it fixes 8.7% of the frozen model's errors, against 17.0% for LSTA on PTB-XL(5.4% vs. 11.8% with 50% contaminated support on PTB-XL full test set; full PTB-XL test set).
  • At equal update size, LSTA's direction is better. Under 50% contamination, LSTA beats direct LayerNorm adaptation rescaled to the same norm by +1.44 [+1.21, +1.66] on PTB-XL.

We thank the reviewer again; the early-window experiment and the entropy measurement made the paper's claims more precise. We hope these answers address the reviewer's concerns, and we would be happy to answer any further questions.

Reviewer_3rHWOpenReview ↗

Thank you for your responses, need more clarification on the motivation

Thank you to the authors for their responses to my points. I appreciate your attention to detail and your direct addressing of my primary points. While your experimental results are convincing and help clear up some points, I am still not certain about the usefulness of this setting in practice or if it represents a sufficiently-wide task to warrant acceptance. I agree with your motivation, but I'm unsure of how these processes are done in practice and how much lift your method gives over existing, simpler approaches. I would appreciate more motivation on the practical findings from your method beyond the narrow tasks on which you test before raising my score.

Comments on the paper

2 comments
Area_Chair_p1yYOpenReview ↗

Discussion period ending soon

Dear Reviewers,

Thank you for your reviews of this submission. The authors have posted responses addressing the concerns raised.

With limited time remaining in the discussion period, please take a moment to review the authors' responses and post a brief follow-up indicating whether your concerns have been addressed or what remains outstanding. If your assessment changes as a result, please update your score accordingly.

Active engagement at this stage is essential for a fair and well-substantiated final decision. Thank you for your time and effort in reviewing this submission.

Best regards, AC

AuthorsOpenReview ↗

Summary of the discussion

Dear Area Chair and Senior Area Chair,

We thank you and all reviewers for your time and for the constructive discussion, which has substantially strengthened our manuscript. The revised paper is uploaded, with all changes highlighted in yellow. We briefly summarize the outcome:

Reviewer 3rHW (2 → 3, weak accept). Concerns on support selection, confident errors, clinical motivation, the choice of LayerNorm (vs. LoRA) and adaptation efficiency were addressed with chronological and later-visit experiments, a confident-error analysis, a LoRA comparison and cost measurements. The reviewer described the implementation as "rigorous" with "robust results across several datasets and tasks."

Reviewer pvTq (2 → 3, weak accept). The main concern, whether robustness comes from the update direction or its magnitude, was resolved by the requested norm-matched control, together with subject-level tests, Frozen-PA, comparisons with MT-Net and Meta-Curvature, and additional backbones. The reviewer confirmed that "the norm-matched comparison addresses my main concern."

Reviewer ADC6 (2 → 3, weak accept). Concerns on significance, the low-rank premise and related meta-learning methods were addressed with subject-level paired tests, a direct test of the low-rank premise, Meta-Curvature and CAVIA-style comparisons, a chronological evaluation and additional backbones. The suggested related work (MBrain, GraphS4mer, CBraMod) is now included.

Reviewer PA5L (2). This review describes a different submission ("Significant-subspace Alignment" for regression) and does not refer to any content of our paper. We hope it will not be weighed in the decision.

Reviewer fC25 (1). We have answered each point raised, including the follow-up questions, with direct evidence for the low-rank hypothesis, controlled comparisons with FiLM and other meta-learned update rules, and the redesigned Figure 2. Our considerations on this review are shared in our confidential comments, which we hope will be taken into account.

Thank you again for your guidance.

Kind regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.