pith. sign in

arxiv: 2606.21329 · v1 · pith:UNV4OTYDnew · submitted 2026-06-19 · 💻 cs.LG

MedTS-TTT: Test-Time Training for Medical Time Series Classification

Mingzhi Chen , Yiyu Gui , Guibo Luo This is my paper

Pith reviewed 2026-06-26 14:27 UTC · model grok-4.3

classification 💻 cs.LG
keywords test-time trainingmedical time seriesEEG classificationECG classificationdistribution shiftself-supervised learningfast adaptationsubject-independent evaluation
0
0 comments X

The pith

MedTS-TTT adapts medical time series models to new subjects via single-step test-time training on unlabeled samples.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Medical time series signals such as EEG and ECG exhibit large subject-level heterogeneity that causes fixed models to fail on unseen individuals. The paper introduces MedTS-TTT, which performs test-time adaptation directly on each new unlabeled test sample without requiring target batches or multiple optimization steps. It does so by building on Closed-Loop Self-Alignment Test-Time Training that creates a token-level self-supervised target and applies one fast-weight update inside the layers for closed-loop alignment. A gated convolutional backbone integrates this adaptation with local dynamic modeling and information-flow control. Experiments on four public EEG and ECG datasets under subject-independent splits show the method ranking first in eleven of twelve metric comparisons against nine baselines.

Core claim

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.

What carries the argument

CLSA-TTT, which constructs a token-level self-supervised target and performs a single-step fast-weight update for intra-layer closed-loop alignment

Load-bearing premise

That the token-level self-supervised target constructed by CLSA-TTT combined with a single-step fast-weight update produces effective adaptation for subject-level shifts in medical time series without needing iterative optimization or target-batch statistics.

What would settle it

On a new subject-independent EEG or ECG dataset, if MedTS-TTT shows no accuracy gain over a non-adaptive baseline or requires more than one update step to reach competitive performance, the central claim would be falsified.

Figures

Figures reproduced from arXiv: 2606.21329 by Guibo Luo, Mingzhi Chen, Yiyu Gui.

Figure 1
Figure 1. Figure 1: The comprehensive architecture of MedTS-TTT. (a) Subject-level distribution shift setting and training/test-time adaptation protocol. (b) Overall architecture of the proposed MedTS-TTT framework. 2.1 Spatiotemporal Tokenizer for Medical Time Series Given a medical time series segment X ∈ R B×C×T , we design a spatiotemporal tokenizer to map it into a token sequence H ∈ R B×L×D. Specifically, we apply chann… view at source ↗
Figure 2
Figure 2. Figure 2: Performance comparison of full MedTS-TTT and ablated variants. Ablation Study. To assess the roles of CLSA-TTT and GCB, we perform ablations on 4 variants: full MedTS-TTT, w/o CLSA-TTT, w/o GCB, and w/ Vanilla TTT. The w/o CLSA-TTT variant removes the CLSA-TTT module from GCB, whereas w/o GCB keeps test-time training but replaces GCB with a standard Transformer FFN block. The w/ Vanilla TTT variant replace… view at source ↗
Figure 3
Figure 3. Figure 3: Distribution-Shift and Feature Alignment under Test-Time Training [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
read the original 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes MedTS-TTT, a test-time training framework for medical time series classification (EEG/ECG) that extends Closed-Loop Self-Alignment Test-Time Training (CLSA-TTT) with a Gated Convolutional Backbone (GCB). CLSA-TTT constructs a token-level self-supervised target and applies a single-step fast-weight update for intra-layer alignment, enabling sample-wise adaptation without iterative optimization or target-batch statistics. On four public datasets with subject-independent splits, it reports 11 top-1 rankings out of 12 evaluations across nine baselines and three metrics.

Significance. If the empirical rankings hold after verification of the adaptation mechanism, the work provides a practical, low-overhead TTT solution for subject-level distribution shifts in clinical time series, avoiding the overhead of iterative inner-loop methods and extra adaptation components common in domain adaptation. Public code release supports reproducibility.

major comments (2)
  1. [Abstract and §3] Abstract and §3 (CLSA-TTT description): the central claim that a single-step fast-weight update on the token-level self-supervised target produces effective adaptation for subject-level shifts is load-bearing for attributing the 11/12 top rankings to the TTT mechanism rather than the GCB backbone or training procedure, yet no analysis of update magnitude, gradient norms, or ablation removing the update step is supplied to confirm meaningful parameter movement.
  2. [Experiments] Experiments section (results table): the reported 11/12 top-1 rankings across 9 baselines and 3 metrics on subject-independent splits lacks error bars, statistical significance tests, or per-subject variance, making it impossible to determine whether the gains are robust or could arise from implementation details of the GCB rather than CLSA-TTT.
minor comments (2)
  1. [§3] Notation for the fast-weight update and token-level target construction should be formalized with equations to allow exact reproduction.
  2. [Experiments] Baseline descriptions are referenced but not detailed in the abstract; a table summarizing each baseline's key hyperparameters would improve clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting the need for stronger evidence on the adaptation mechanism and result robustness. We address each major comment below and will incorporate revisions to address the concerns.

read point-by-point responses

  1. Referee: [Abstract and §3] Abstract and §3 (CLSA-TTT description): the central claim that a single-step fast-weight update on the token-level self-supervised target produces effective adaptation for subject-level shifts is load-bearing for attributing the 11/12 top rankings to the TTT mechanism rather than the GCB backbone or training procedure, yet no analysis of update magnitude, gradient norms, or ablation removing the update step is supplied to confirm meaningful parameter movement.

    Authors: We agree that the manuscript would benefit from explicit verification that the single-step update induces meaningful adaptation. In the revision we will add: (i) statistics on update magnitudes and gradient norms across layers and samples, and (ii) an ablation that disables the fast-weight update while retaining the GCB backbone and training procedure, allowing direct attribution of gains to CLSA-TTT. revision: yes

  2. Referee: [Experiments] Experiments section (results table): the reported 11/12 top-1 rankings across 9 baselines and 3 metrics on subject-independent splits lacks error bars, statistical significance tests, or per-subject variance, making it impossible to determine whether the gains are robust or could arise from implementation details of the GCB rather than CLSA-TTT.

    Authors: We acknowledge that the current table does not report variability or significance. In the revised manuscript we will rerun all experiments with multiple random seeds, report mean ± standard deviation, and include statistical significance tests (e.g., paired t-tests) against the strongest baselines. Per-subject performance breakdowns will also be added to the supplement. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical extension of prior TTT without reductive derivations

full rationale

The paper presents MedTS-TTT as a practical extension of existing Test-Time Training methods (specifically CLSA-TTT) combined with a Gated Convolutional Backbone for medical time series. No equations, derivations, or first-principles predictions appear in the provided text. Central claims rest on experimental top-1 rankings across public datasets and baselines rather than any fitted parameter or self-citation that reduces the result to its inputs by construction. The work is self-contained as an applied method with external empirical validation.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only abstract available; no explicit free parameters, axioms, or invented entities identifiable. The method implicitly assumes the self-supervised alignment target is valid for medical signals.

pith-pipeline@v0.9.1-grok · 5798 in / 1100 out tokens · 15252 ms · 2026-06-26T14:27:24.763183+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

26 extracted references

  1. [1]

    Biomedizinische Technik/Biomedical Engineering40(Ergänzungsband 1), 317–318 (1995)

    Bousseljot, R.D., Kreiseler, D., Schnabel, A.: Nutzung der EKG-signaldatenbank CARDIODAT der PTB über das internet. Biomedizinische Technik/Biomedical Engineering40(Ergänzungsband 1), 317–318 (1995)

    1995

  2. [2]

    IEEE Journal of Biomedical and Health Informatics27(11), 5216–5224 (2023)

    Chang, H., Liu, B., Zong, Y., Lu, C., Wang, X.: EEG-based Parkinson’s disease recognition via attention-based sparse graph convolutional neural network. IEEE Journal of Biomedical and Health Informatics27(11), 5216–5224 (2023)

    2023

  3. [3]

    Physiological Measurement27(11), 1091–1106 (2006)

    Escudero,J.,Abásolo,D.,Hornero,R.,Espino,P.,López,M.:Analysisofelectroen- cephalograms inAlzheimer’s disease patients with multiscale entropy. Physiological Measurement27(11), 1091–1106 (2006)

    2006

  4. [4]

    In: Proceedings of the ACM on Web Conference 2025

    Fan,W.,Fei,J.,Guo,D.,Yi,K.,Song,X.,Xiang,H.,Ye,H.,Li,M.:Towardsmulti- resolution spatiotemporal graph learning for medical time series classification. In: Proceedings of the ACM on Web Conference 2025. pp. 5054–5064. ACM (2025)

    2025

  5. [5]

    In: Advances in Neural Information Processing Systems

    Gandelsman, Y., Sun, Y., Chen, X., Efros, A.A.: Test-time training with masked autoencoders. In: Advances in Neural Information Processing Systems. vol. 35, pp. 29374–29385 (2022)

    2022

  6. [6]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Han, D., Li, Y., Li, T., Cao, Z., Wang, Z., Song, J., Cheng, Y., Zheng, B., Huang, G.: ViT3: Unlocking test-time training in vision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 51–61 (2026), oral presentation

    2026

  7. [7]

    Nature Medicine25(1), 65–69 (2019)

    Hannun, A.Y., Rajpurkar, P., Haghpanahi, M., Tison, G.H., Bourn, C., Turakhia, M.P., Ng, A.Y.: Cardiologist-level arrhythmia detection and classification in am- bulatory electrocardiograms using a deep neural network. Nature Medicine25(1), 65–69 (2019)

    2019

  8. [8]

    Physical and Engineering Sciences in Medicine47(3), 1037–1050 (2024)

    Kachare, P., Puri, D., Sangle, S.B., Al-Shourbaji, I., Jabbari, A., Kirner, R., Alameen, A., Migdady, H., Abualigah, L.: Lcadnet: a novel light CNN architecture for EEG-based Alzheimer disease detection. Physical and Engineering Sciences in Medicine47(3), 1037–1050 (2024)

    2024

  9. [9]

    Scientific Reports15(1), 7817 (2025)

    Kim, D., Lee, K.R., Lim, D.S., Lee, K.H., Lee, J.S., Kim, D.Y., Sohn, C.B.: A novel hybrid CNN-transformer model for arrhythmia detection without R-peak identification using Stockwell transform. Scientific Reports15(1), 7817 (2025)

    2025

  10. [10]

    Biomedical Signal Processing and Control91, 105872 (2024) 10 M

    Li, J., Li, X., Mao, Y., Yao, J., Gao, J., Liu, X.: Classification of Parkinson’s disease EEG signals using 2D-MDAGTS model and multi-scale fuzzy entropy. Biomedical Signal Processing and Control91, 105872 (2024) 10 M. Chen et al

    2024

  11. [11]

    In: International Conference on Learning Representations (2024)

    Liu, Y., Hu, T., Zhang, H., Wu, H., Wang, S., Ma, L., Long, M.: iTransformer: Inverted transformers are effective for time series forecasting. In: International Conference on Learning Representations (2024)

    2024

  12. [12]

    Miltiadous, A., Tzimourta, K.D., Afrantou, T., Ioannidis, P., Grigoriadis, N., Tsa- likakis, D.G., Angelidis, P., Tsipouras, M.G., Glavas, E., Giannakeas, N., Tzallas, A.T.: A dataset of EEG recordings from: Alzheimer’s disease, Frontotemporal de- mentia and healthy subjects (2024)

    2024

  13. [13]

    In: International Conference on Learning Representations (2023)

    Nie, Y., Nguyen, N.H., Sinthong, P., Kalagnanam, J.: A time series is worth 64 words: Long-term forecasting with transformers. In: International Conference on Learning Representations (2023)

    2023

  14. [14]

    Proceedings of Machine Learning Research, vol

    Sun, Y., Li, X., Dalal, K., Xu, J., Vikram, A., Zhang, G., Dubois, Y., Chen, X., Wang, X., Koyejo, S., Hashimoto, T., Guestrin, C.: Learning to (learn at test time): RNNswithexpressivehiddenstates.In:Proceedingsofthe42ndInternationalCon- ference on Machine Learning. Proceedings of Machine Learning Research, vol. 267, pp. 57503–57522. PMLR (2025)

    2025

  15. [15]

    In: 2020 42nd Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC)

    Tadesse, G.A., Javed, H., Weldemariam, K., Zhu, T.: A spectral-longitudinal model for detection of heart attack from 12-lead electrocardiogram waveforms. In: 2020 42nd Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). pp. 6009–6012 (2020)

    2020

  16. [16]

    In: Advances in Neural Information Processing Systems

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems. vol. 30, pp. 5998–6008 (2017)

    2017

  17. [17]

    Scientific Data7(1), 154 (2020)

    Wagner, P., Strodthoff, N., Bousseljot, R.D., Kreiseler, D., Lunze, F.I., Samek, W., Schaeffter, T.: PTB-XL, a large publicly available electrocardiography dataset. Scientific Data7(1), 154 (2020)

    2020

  18. [18]

    In: Advances in Neural Information Processing Systems

    Wang, Y., Huang, N., Li, T., Yan, Y., Zhang, X.: Medformer: A multi-granularity patching transformer for medical time-series classification. In: Advances in Neural Information Processing Systems. vol. 37, pp. 36314–36341 (2024)

    2024

  19. [19]

    In: Advances in Neural Informa- tion Processing Systems

    Wu, H., Xu, J., Wang, J., Long, M.: Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. In: Advances in Neural Informa- tion Processing Systems. vol. 34, pp. 22419–22430 (2021)

    2021

  20. [20]

    Proceedings of the AAAI Conference on Artificial Intelligence40(33), 27791–27799 (2026)

    Ye, J., Zhang, W., Li, Z., Li, J., Tsung, F.: Medspaformer: A transferable trans- former with multi-granularity token sparsification for medical time series classi- fication. Proceedings of the AAAI Conference on Artificial Intelligence40(33), 27791–27799 (2026)

    2026

  21. [21]

    Applied Sciences9(16), 3328 (2019)

    Zhang, Y., Li, J.: Application of heartbeat-attention mechanism for detection of myocardial infarction using 12-lead ECG records. Applied Sciences9(16), 3328 (2019)

    2019

  22. [22]

    In: International Conference on Learning Representations (2023)

    Zhang, Y., Yan, J.: Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting. In: International Conference on Learning Representations (2023)

    2023

  23. [23]

    Proceedings of the AAAI Conference on Artificial Intelligence35(12), 11106–11115 (2021)

    Zhou, H., Zhang, S., Peng, J., Zhang, S., Li, J., Xiong, H., Zhang, W.: Informer: Beyond efficient transformer for long sequence time-series forecasting. Proceedings of the AAAI Conference on Artificial Intelligence35(12), 11106–11115 (2021)

    2021

  24. [24]

    In: Proceedings of the 39th International Conference on Machine Learning

    Zhou, T., Ma, Z., Wen, Q., Wang, X., Sun, L., Jin, R.: FEDformer: Frequency enhanced decomposed transformer for long-term series forecasting. In: Proceedings of the 39th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 162, pp. 27268–27286. PMLR (2022)

    2022

  25. [25]

    Proceedings of the AAAI Conference on Artificial Intelligence39(13), 14529–14537 (2025) MedTS-TTT: Test-Time Training for Medical Time Series Classification 11

    Zhou, Y., Zhao, S., Wang, J., Jiang, H., Li, S., Luo, B., Li, T., Pan, G.: Personalized sleep staging leveraging source-free unsupervised domain adaptation. Proceedings of the AAAI Conference on Artificial Intelligence39(13), 14529–14537 (2025) MedTS-TTT: Test-Time Training for Medical Time Series Classification 11

    2025

  26. [26]

    Biomedical Signal Processing and Control114, 109321 (2026)

    Zini, S., Barbera, T., Bianco, S., Napoletano, P.: Alzheimer’s disease classification from EEG using a multiscale temporal deep network. Biomedical Signal Processing and Control114, 109321 (2026)

    2026