List of Papers Browse by Subject Areas Author List
Abstract
Medical time series (MedTS) signals such as electroencephalography (EEG) and electrocardiography (ECG) support many clinical applications. However, substantial subject-level heterogeneity often induces subject-level distribution shift, causing a fixed parameter set to generalize poorly to unseen individuals. Compared with domain adaptation methods that often depend on extra adaptation components or target-batch statistics, Test-Time Training (TTT) provides a more practical solution for sequential clinical data by enabling online adaptation from unlabeled test samples. However, many representative TTT methods require iterative inner-loop optimization, increasing test-time overhead. In this paper, we propose MedTS-TTT, a test-time training framework for medical time series modeling. MedTS-TTT is built upon Closed-Loop Self-Alignment Test-Time Training (CLSA-TTT) and a Gated Convolutional Backbone (GCB). CLSA-TTT constructs a token-level self-supervised target and performs a single-step fast-weight update for intra-layer closed-loop alignment, enabling rapid sample-wise adaptation without iterative inner-loop optimization. GCB combines CLSA-TTT-based fast adaptation and token-level fusion with a gated convolutional branch to balance local dynamic modeling and information-flow control. On 4 public datasets (2 EEG and 2 ECG) with subject-independent splits, MedTS-TTT achieves 11 top-1 rankings out of 12 evaluations across 9 baselines and 3 metrics. The code is publicly available at https://github.com/mingzhi-c/MedTS-TTT.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5029_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/mingzhi-c/MedTS-TTT
Link to the Dataset(s)
APAVA: https://osf.io/jbysn/
ADFTD: https://openneuro.org/datasets/ds004504
PTB: https://physionet.org/content/ptbdb/1.0.0/
PTB-XL: https://physionet.org/content/ptb-xl/1.0.3/
BibTex
@InProceedings{CheMin_MedTSTTT_MICCAI2026,
author = { Chen, Mingzhi AND Gui, Yiyu AND Luo, Guibo},
title = { { MedTS-TTT: Test-Time Training for Medical Time Series Classification } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes MedTS-TTT, an online test-time training framework for medical time series classification that addresses subject-level distribution shifts by utilizing a Closed-Loop Self-Alignment (CLSA) mechanism for single-step fast-weight updates, integrated within a Gated Convolutional Backbone to ensure low-latency adaptation without target-domain labels.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The study addresses a relevant clinical challenge regarding the performance degradation of medical time series models due to inter-subject heterogeneity, motivating online test-time training as a feasible approach to bypass the need for target-domain batches. 2.The Closed-Loop Self-Alignment Test-Time Training (CLSA-TTT) mechanism offers an alternative to traditional iterative inner-loop optimization by employing a single-step fast-weight update, which is a practical consideration for latency-sensitive clinical inference applications. 3.The experimental setup utilizes subject-independent splits across four distinct clinical datasets encompassing both EEG and ECG signals, providing empirical support for the method’s potential generalizability across these different medical domains.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The theoretical rationale for the specific token-level self-supervised objective could be further elaborated; explicitly discussing or empirically demonstrating why aligning the normalized key-value residual to the fast-parameter transformed keys is better suited for clinical time-series shifts compared to established contrastive or predictive test-time objectives would strengthen the methodological foundation. 2.The experimental evaluation would be more comprehensive if it incorporated comparisons with contemporary test-time adaptation (TTA) frameworks; while the current baselines effectively demonstrate improvements over static time-series backbones and MedTS-specific models, including recent TTT methods tailored for sequential data would better contextualize the proposed adaptation mechanism’s relative advantages.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper proposes MedTS-TTT, a single-step test-time training framework with a CLSA-TTT mechanism, to address distribution shifts caused by inter-subject heterogeneity in medical time-series. It achieves low-latency, label-free online adaptation, demonstrating robust cross-subject generalization across four EEG/ECG datasets through rigorous, clinically relevant experiments. Although theoretical justifications for the self-supervised objectives and quantitative comparisons with recent sequential TTA baselines could be expanded, the rapid weight update mechanism effectively balances adaptation capability with computational latency. Given its clear methodological and practical value, the paper meets the acceptance criteria for MICCAI.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The article introduces MedTS-TTT, which is a test-time training approach for medical time series classification in order to tackle the problem of subject-level distribution shift in EEG and ECG data by means of a two-component design: (1) the Closed-Loop Self-Alignment TTT (CLSA-TTT) algorithm, which is a token-level self-supervision method for model fine-tuning during inference time via one-shot fast weight updating in order to avoid repeated inner-loop optimization iterations, and (2) the Gated Convolutional Backbone (GCB) approach that combines depth-wise convolutions, gating, and CLSA-TTT into a unified sequence modeling layer. The proposed framework is benchmarked on 4 publicly available datasets with subject-independent splits and compared to 9 baseline methods in terms of three different performance metrics, with the results of obtaining 11 out of 12 best-1 scores reported. The clinical problem statement and efficiency improvement provided over vanilla TTT approach is convincing enough. However, the key innovation of CLSA-TTT is incremental, since the fast weights method introduced in Sun et al. [14] only differs from the latter in one gradient step iteration.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- Clinical motivation behind the problem: distribution shift at a subject level is a true limitation to implementing EEG/ECG models in practice and the annotation-free setup using online adaptation is quite realistic.
- The efficiency improvement is promising: reducing p50/p95 latency by half and doubling throughput in comparison with TTT alone resolves an existing practical issue.
- Empirical investigation of a wide scope: 4 data sets, 9 baselines, and 3 metrics with split evaluation per subject ensure fair and replicable results.
- An ablation study is well-thought-out, clearly distinguishing the effect of CLSA-TTT from GCB.
- Spectral entropy shift and t-SNE alignment explains the rationale behind the adaptation process.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- The CLSA-TTT method basically uses the same mechanism as TTT-RNN hidden states in reference [14] with its contribution being only in making the SGD a one-step optimization without any explanation as to why just a single step was enough for MedTS.
- Lack of comparisons to domain adaptation methods (i.e., TTT++ or TENT) and test-time normalization techniques like BN adaptation that would naturally compare against the problem setup.
- High standard deviation on ECG-PTB performance (±3.12% accuracy and ±4.81% F1-score) makes the improvements over MedGNN non-statistically significant as they are in error margins.
- Only CLSA-TTT vs. vanilla TTT is compared. Training wall-clock time, parameter count, and FLOPs relative to baselines are not reported, makes the computational overhead analysis is narrow.
- Lack of theoretical grounding for using LN(Z) ≈ V−K as self-supervised learning criteria. It appears to be taken directly from reference [14] without any justification.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
- The PTB results show large standard deviations (e.g., ±4.81% F1). How many random seeds were used? Is the gain over MedGNN (+0.31% F1) statistically significant, and can you provide p-values or confidence intervals?
- Why is the self-supervised target V−K chosen? It would be helpful if there were an intuitive or theoretical reason for why alignment of Z with the key-value residual is useful for medical time-series data. -What is the total parameter count and FLOPs of MedTS-TTT compared to MedFormer and MedGNN? A fairer efficiency comparison would include all baselines, not only vanilla TTT.
- how the head-wise adaptive step size η in Eq. (4) is it set or learned? Is it sensitive to initialization, and does it require separate tuning per dataset?
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a well-motivated problem of distribution shift at the subject level in clinical time-series data and presents an effective and efficient solution with a comprehensive evaluation showing positive results consistently across 4 datasets, 9 baselines, and 3 metrics. However, there are several reservations about the paper. First, the main methodological novelty (CLSA-TTT) is incremental compared to Sun et al. [14] since it is not clear whether a single update step is sufficient for MedTS adaptation without any theoretical motivation. Second, the lack of adaptation methods at test-time, like TENT or TTT++, makes it hard to evaluate the gains from the CLSA-TTT approach. Third, the high variance of ECG-PTB (±4.81% F1) makes the advantage over MedGNN marginal without statistical significance checks. Fourth, the negative AUROC on PTB compared to MedGNN is never addressed.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors addressed all reviewers’ comments and it is clear now.
Review #3
- Please describe the contribution of the paper
The authors propose MGK-DTI, a dual-model neural network framework for predicting drug–target interactions using raw drug molecule and target protein information. The main contributions include the introduction of a Mambaformer-based sequence branch to capture multi-scale dependencies, the development of an HGGCN-based structural branch for topological feature extraction, and the incorporation of KAN to enhance nonlinear interaction modeling. Comprehensive experiments demonstrate that MGK-DTI outperforms state-of-the-art methods in DTI prediction.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper presents several notable strengths. First, MedTS-TTT employs a coherent series of transformations to project the original input into a structured embedding space H ̃, forming a logically organized and end-to-end transformation pipeline. Furthermore, the integration of CLSA-TTT with the gated fusion module offers a principled and stable solution for domain adaptation, enabling efficient cross-subject generalization without iterative optimization. Most importantly, the paper conducts comprehensive evaluations across four public datasets using multiple performance metrics, complemented by thorough ablation studies, which collectively substantiate the model’s robustness and generalizability.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
Training (CLSA-TTT) module, whose effectiveness is demonstrated through the ablation study. However, the design of the TTT loss function lacks rigorous mathematical justification, and the authors provide no theoretical motivation for the chosen formulation. The current explanation reads more as a post-hoc description than a principled derivation, leaving open the question of whether the design reflects systematic insight or an empirically successful but poorly understood choice. Furthermore, while MedTS-TTT outperforms state-of-the-art baselines, the margin of improvement — often only a few tenths of a percentage point — is not accompanied by any discussion of its clinical or practical significance, making it difficult to assess whether the gains are meaningful in real-world medical time series analysis. The reported results also reveal that MedTS-TTT exhibits notably higher variance compared to methods such as Medformer and MedGNN, raising concerns about the method’s stability and reliability, which are critical considerations in clinical applications. Finally, the integration of TTT introduces a fundamental comparison fairness concern. Since TTT enables the model to update its parameters using each test sample — even in a self-supervised manner — the model effectively gains access to test-time information that competing baselines do not. This mechanism resembles transductive or meta-learning settings more than standard generalization benchmarks, and the paper does not adequately address or control for this disparity when making comparisons.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a novel yet effective method for medical time series classification. However, the method lacks rigorous mathematical support, and the comparisons contain gaps that need to be further addressed.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank the reviewers and respond to the main concerns below.
Global response: R1-W1&R2-W5,Q2&R3-W1: TTT loss/V-K rationale. The objective targets current-sample adaptation, not raw-signal reconstruction. K is a stable token anchor, V carries transformed content, and V-K captures residual current temporal patterns. Aligning LN(KW_fast) with V-K trains fast weights to model this sample-specific residual; normalization reduces sensitivity to EEG/ECG amplitude/scale shifts across subjects. Applying the updated W_fast to Q closes the correction loop. We will add this motivation and link to clinical shifts.
R1-W2&R2-W2: TTA/DA comparisons. As stated in the Introduction, our goal is to improve TTT for sequential MedTS inference without labels, target-domain training, future samples, or reliable target-batch statistics. The closest controlled baseline is same-backbone vanilla TTT, which shares per-sample self-supervised adaptation. Many DA/TENT-style methods assume a target domain/batch, extra modules, or entropy/statistics conditions that do not match single-sample deployment. As described in Sec. 2.2, CLSA-TTT makes one closed-loop fast-weight update inside the sequence layer, avoiding iterative inner-loop optimization.
R2-W3,Q1&R3-W3: PTB variance/significance. We agree PTB is variable. Similar variance appears in strong baselines, e.g., Crossformer (±3.79 ACC, ±7.19 F1) and FEDformer (±2.54 ACC, ±4.37 F1), suggesting dataset/protocol instability rather than a method-specific issue. The PTB F1 gain over MedGNN is modest and PTB AUROC is lower, so our claim rests on the 4-dataset/3-metric pattern: 11/12 top-1 results and average Accuracy/F1/AUROC gains of 0.90/0.65/0.83 over the strongest baseline.
Reviewer 2: R2-W1: novelty/one-step design. CLSA-TTT is not vanilla TTT with fewer iterations. It adds an in-layer closed-loop fast-weight mechanism: each test sample forms a residual target, updates a small fast-weight matrix once, and immediately applies it to the query stream for sample-adaptive correction. This is embedded inside the sequence layer rather than used as an external optimizer. The one step targets low-latency clinical inference. Replacing CLSA-TTT with vanilla TTT lowers average accuracy by 2.64%.
R2-W4,Q3: computational overhead, parameters, FLOPs. We compare adaptation rules under the same backbone. Versus vanilla TTT, CLSA-TTT reduces p50 latency from 29.35 to 14.65 ms and p95 from 29.70 to 14.92 ms, while increasing throughput from 4374.23 to 8744.17 samples/s. Only a small fast-weight matrix is updated at test time; the slow backbone is shared.
R2-Q4: head-wise step size. The step size η is learned during source-domain training as a bounded [0,1] parameter, then fixed at test time. It needs no target labels, per-subject tuning, or dataset-specific test-time adjustment, and was stable in practice.
Reviewer 3: R3-W2: practical significance. MedTS-TTT is not only a small static accuracy gain: it adapts online to unseen subjects without labels or target-domain data, matching EEG/ECG inference under subject shift, and is faster than vanilla TTT. Its value is label-free online adaptability with improved efficiency.
R3-W4: fairness of using test-time samples. Our evaluation is explicitly a test-time training setting, not static inference. The model only uses the current unlabeled sample to update a small set of fast parameters; it does not use labels, target-domain data, or future test samples. We include w/o CLSA-TTT and w/ Vanilla TTT as controls. The full model improves average accuracy by 4.05% over w/o CLSA-TTT and by 2.64% over vanilla TTT, showing that the gain is not merely from accessing the test input.
Possible review mismatch: R3’s contribution summary appears to describe an unrelated DTI paper (MGK-DTI), while the remaining comments discuss MedTS-TTT. We hope the above clarifications address the substantive concerns.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The reviewers have mixed opinions with ratings as SA, WA, and WR. The reviewers appreciated the well-motivated problem of subject-level distribution shift, and the interesting test-time training approach with the closed-loop self-alignment and gated convolutional backbone. The authors are encouraged to clarify the following concerns raised by the reviewers: incremental novelty compared to Sun et al, lack of comparison with domain-adaptive methods, narrow computational overhead analysis, marginal improvements over the baselines, etc.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Authors clarified the main concerns in the rebuttal. There is agreement among reviewers regarding acceptance.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper addresses the clinically relevant problem of subject-level distribution shifts in medical time series using a novel Closed-Loop Self-Alignment (CLSA) mechanism for efficient, low-latency online adaptation. In their rebuttal, the authors effectively resolved initial reviewer concerns by clarifying the theoretical rationale for the self-supervised objective, justifying their choice of test-time training baselines, and explaining that the variance on the ECG-PTB dataset aligns with established benchmarks. Because all reviewers found the rebuttal convincing and upgraded their final recommendations to “Accept,” the paper meets the high standards for this conference.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
After considering the reviews and rebuttal, I lean toward acceptance. The paper addresses an important and practically relevant problem of subject-level distribution shift in medical time-series analysis and proposes a test-time training framework tailored to single-sample adaptation without labels or target-domain data. The rebuttal provides additional clarification regarding the closed-loop self-alignment mechanism, demonstrates consistent improvements over both the non-adaptive backbone and vanilla TTT, and presents evidence that the proposed approach achieves substantially lower latency than standard TTT while maintaining competitive performance. The authors also clarify the intended deployment setting and explain why many domain-adaptation methods are not directly comparable.
I agree that concerns remain regarding the degree of novelty relative to prior TTT methods, the limited comparison with broader adaptation approaches, and the modest gains on some datasets. However, these issues do not appear to undermine the technical validity of the method.
Given the practical motivation, generally consistent empirical results, and favorable efficiency-performance tradeoff, I believe the paper makes a worthwhile contribution and recommend Acceptance.