Pith. sign in

REVIEW 4 major objections 5 minor 48 references

Discrepancy-Aware Contrastive Adaptation in Medical Time Series Analysis

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read DAAC combines an autoencoder reconstruction-error feature with an attention-based multi-view contrastive learner and reports beating all seven baselines on EEG/ECG diagnosis of Alzheimer's, Parkinson's, and myocardial infarction, including

desk verdict The method is a plausible new combination, but the tables contradict the central SOTA claim—not reviewable as is. read the letter →

arxiv 2508.05572 v1 pith:2P4YIXWA submitted 2025-08-07 cs.HC

Yifan Wang , Hongfeng Ai , Ruiqi Li , Maowei Jiang , Ruiyuan Kang , Jiahua Dong , Cheng Jiang , Chenzhong Li This is my paper
classification cs.HC
keywords medicaltimeseriescontrastivelearningmulti-viewrepresentationdiscrepancyestimationdiseasediagnosisEEGECGdomainshift
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

DAAC's central claim is that two mechanisms together — a cheap disease-probability feature and a contrastive learner that adaptively discovers its own views — remove the two main bottlenecks in medical time-series diagnosis: expensive labels and handcrafted positive/negative pairs. The disease-probability feature is the reconstruction error of an AE-GAN autoencoder trained on external healthy data, appended to every sample as one extra input dimension; the Learnable Multi-views Contrastive Framework (LMCF) uses multi-head attention to generate view-level representations and contrasts them within and between views, so no manual pairing is needed. The paper reports that DAAC beats all seven baselines on three public datasets (Alzheimer's EEG, Parkinson's EEG, myocardial infarction ECG) across six metrics, and that most of the advantage survives when only 10% of labels are used for fine-tuning. If the claim is right, hospitals with small labeled cohorts can lean on publicly available healthy recordings to train diagnostic models without overfitting and without annotation bias.

What carries the argument

Two coupled modules carry the argument. The Discrepancy Estimator is an encoder–decoder generator (with a GAN discriminator) pretrained only on external normal data; its sequence-level reconstruction error $E$ is appended to every target sample (Eq. 3), which is the entire channel by which external knowledge enters the model — the magnitude of $E$ is read as the probability of disease. The Adaptive Contrastive Learner is a dilated-convolution encoder plus a multi-head attention (MHA) block that produces $V$ view-level representations per sample; an inter-view contrastive loss makes views distinguishable while an intra-view loss separates subjects within each view, and four hierarchical InfoN

What would settle it

Two concrete checks. (1) Replace $E$ in Eq. (3) with a random-noise feature of the same shape and re-run fine-tuning on all three datasets; if DAAC's AUROC does not fall, the discrepancy feature is not carrying the gain. (2) Measure the reconstruction error's stand-alone AUROC as a disease classifier on a target set whose acquisition differs from the external data — the paper reports 0.974 for the in-domain AD validation but not the cross-center value; if that score collapses toward 0.5 while DAAC still beats its no-DE ablation, the reported advantage comes from the contrastive learner, not fr

Watch

Extended reading notes

Core claim

The discovery the paper is trying to establish is that a sample's deviation from a learned normal distribution — expressed as a single sequence-level reconstruction error — can be turned into a useful input feature for medical diagnosis, and that multi-head attention can replace manual contrastive-pair design. DAAC first trains an encoder–decoder generator with a GAN discriminator on external data from healthy subjects; at inference the reconstruction error $E = \mathrm{MSE}(G_{gen}(\bar{x}_i), \bar{x}_i)$ is concatenated with the raw signal (Eq. 3), making 'abnormal' a scalar feature. The augmented data feeds an encoder with a dilated-convolution backbone and a multi-head attention block th

Load-bearing premise

The load-bearing premise, stated in Section 4.1.3 (Eqs. 2–3), is that the reconstruction error produced by an autoencoder trained only on external healthy data indirectly reflects the abnormality of a target sample — that is, its probability of disease — so one scalar can be fused into every training sample; the paper's own Appendix C3 shows the premise is fragile, because on the AD task the external and target cohorts differ (mean age 63.6 vs 72.5) and adding the discrepancy

Editorial extensions

If this is right

  • If the reported margins hold, any center with a small labeled cohort and access to a public pool of healthy recordings can reproduce the full pipeline without per-disease engineering: the same DE + ACL recipe is demonstrated on EEG and ECG across three diseases.
  • The 10%-label results (e.g., AD AUROC 98.10, TDBrain AUROC 95.27, PTB AUROC 96.30 under full fine-tuning) imply that the stage-2 representation carries most of the diagnostic information, which is what makes low-resource deployment plausible.
  • The ablation study attributes an average gain of roughly 3.1% in F1 to the view-contrastive loss, so the attention-based view mechanism — not the discrepancy feature alone — is the main engine of the reported improvement.
  • Appendix C3 documents that adding the DE feature alone to COMET lowers AD AUROC from 94.44 to 93.97 under cross-center shift, so the discrepancy feature is a complement rather than a standalone signal; the full two-module system is the deployable unit.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Not tested in the paper: whether the discrepancy feature $E$ would work better as a gating signal or a loss weight than as one concatenated scalar; the mutual-information analysis ($E$ at 0.0295 vs a 0.0159 ceiling for raw features) suggests the score carries signal, but a single scalar is a coarse way to spend it.
  • The authors leave implicit that the method's cross-center gain should shrink as the demographic and protocol gap between external and target data grows; the AD case (external mean age 63.6 vs 72.5) is a mild example of exactly the shift that would test this.
  • Editorial flag: Appendix D.2 promises a full loss-weight sensitivity table but cites it as 'Table ??', which is not present in this version, so the robustness claim for the fixed 1:1:1:1:2 weighting currently rests on the prose summary alone.
  • A natural next step the paper names as future work is joint optimization of the five loss weights; the saturation visible when the view weight grows from 2 to 3 suggests per-dataset weight adaptation could sharpen the 10%-label regime, where DAAC's F1 edge over the no-DE ablation is thinnest (e.g., PTB: 85.61 vs 85.35).
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes DAAC, a three-stage framework for medical time-series diagnosis (AD, Parkinson's disease/TDBrain, myocardial infarction/PTB). Stage 1 trains an AE-GAN on external normal data and uses the reconstruction error as an extra 'discrepancy' feature; Stage 2 learns representations via hierarchical contrastive losses augmented with an MHA-based inter/intra-view contrastive loss; Stage 3 fine-tunes for classification. The abstract and Section 5.1 claim consistent superiority over seven baselines, including under 10% labeled data. The manuscript includes ablation studies, hyperparameter sensitivity, domain-shift analysis, and efficiency reporting.

Significance. If the empirical claims were sound, the combination of a reconstruction-based anomaly feature with an adaptive multi-view contrastive learner would be a useful contribution to low-label medical time-series diagnosis. The paper also reports extensive ablations (loss weights, external-data ratio, discrepancy-feature design) and reproducibility details in the appendices, which are valuable. However, the central claim is purely empirical and the presented evidence is internally inconsistent and contradicted by the paper's own tables. As submitted, the results do not establish the claimed state-of-the-art performance, and the empirical protocol has a selection-on-test risk. The ideas are worth exploring, but the manuscript's current numbers need substantial rework and likely re-experimentation.

major comments (4)
  1. [Table 2 vs Appendix D/E] The central performance claim is not supported by internally inconsistent tables. For the AD 100% configuration 1,1,1,1,0 (i.e., P+R+S+O / COMET), Table 2 reports AUROC 94.44 and AUPRC 94.43, while Table D1 reports AUROC 93.60 and AUPRC 94.10, despite identical Accuracy (84.50), Precision (88.31), Recall (82.95), and F1 (83.33). Likewise, for 1,1,1,1,2 (COMET+ACL), Table 2 gives AUROC/AUPRC 96.22/96.03, but Table D1 gives 96.40/96.30. Table E2's P+R+S+O+V for PTB 100% reports Accuracy 89.57 and AUROC 95.70, whereas Table 2's DAAC reports Accuracy 93.65 and AUROC 95.56 for the same configuration. These are not minor typos: they affect the very numbers used to claim superiority. The reader cannot tell which table reflects the true experiment, so the central empirical claim is unsubstantiated as written.
  2. [Section 5.1 / Table 1] The statement 'Our proposed DAAC model excels in all metrics' is contradicted by Table 1. COMET+DE has higher Precision (80.41 vs 80.34), AUROC (86.32 vs 85.96), and AUPRC (85.85 vs 85.40) than DAAC. In Table 2, AD 10%, COMET+DE also exceeds DAAC on AUROC (98.27 vs 98.10) and AUPRC (98.30 vs 98.19). Moreover, no significance tests are reported; in several cases the differences are within one standard deviation (e.g., Table 1 DAAC vs COMET+DE). Thus the abstract's 'consistently outperforms other seven baselines' is not established by the data shown.
  3. [Eq. (12), Appendix D.2] The final loss weights λS:λR:λE:λT:λV = 1:1:1:1:2 appear to be selected on the same target datasets used for the headline results. Appendix D.2 reports a sensitivity analysis on AD, PTB, and TDBrain at both 100% and 10% label rates (Table D1), and the chosen weight V=2 corresponds to the best AUROC in several settings (e.g., AD 100%: 96.40). This is a selection-on-test procedure unless a separate validation split was used, which is not described. Section E.2 simultaneously claims 'We did not tune these weights,' which is inconsistent with D.2. The magnitude of the reported gains is therefore not a fair assessment of an a priori configuration.
  4. [Section 4.1.3 / C3] The paper treats the reconstruction error E as 'the probability of disease' and fuses it into every training sample. This assumption is load-bearing for the discrepancy-aware design, but Appendix C3 documents that adding DE to COMET decreases AD AUROC from 94.44 to 93.97, showing that the reconstruction error can be misleading under the exact cross-center shift the method targets. The validation offered (KDE in F.2 and MI of 0.0295 in E.1) is weak and does not establish that E is an informative monotone disease-probability signal in the target domain. If E is uninformative or noisy, DAAC reduces to ACL; the paper does not provide evidence that ACL alone does not drive the results.
minor comments (5)
  1. [Abstract / Section 5] The source code link is a placeholder ('xxxxx'), despite the abstract stating 'We release the source code.' Please provide a working repository or remove the claim.
  2. [Section 3-4 headings] The section numbering is duplicated: '3 Proposed Method:DAAC' is immediately followed by '4 Proposed Method:DAAC'. This should be corrected.
  3. [Throughout] There are numerous typos and inconsistent terms: 'Fine Tuninig', 'Respecively', 'PDB dataset' (should be PTB), 'DaulMultiHeadTSEncoder', and the header 'Appendix G.2 Loss Weight Sensitivity Analysis' appearing inside Appendix D.2. Please proofread carefully.
  4. [Table E2] The AD 10% 'P+R+S+O' row contains what appear to be misplaced values (e.g., '92.52±2.36' inserted under F1/AUROC/AUPRC). The PTB 100% 'P+R+S+O+V' row does not match Table 2's DAAC row. These need to be reconciled or explicitly explained.
  5. [References / Appendix X] In Section 4, 'Appendix X demonstrates this experiment in detail' should refer to the actual appendix letter (C3 or F) rather than a placeholder.

Circularity Check

1 steps flagged · score 4.0 of 10

No equation-level circularity, but the reported SOTA is partly constructed by loss-weight selection on the same target datasets; the manuscript also contains internal metric inconsistencies.

  1. fitted input called prediction [Section 4.2 (Eq. 12), Appendix D.2 (Table D1), Abstract]
    "λS : λR : λE : λT : λV = 1 : 1 : 1 : 1 : 2in our practice. The details about loss weight sensitivity could be seen in Appendix D.2 / ... we performed a comprehensive sensitivity analysis on the weighting strategy... evaluated the model on three datasets (AD, PTB, TDBrain) under both 100% and 10% data regimes... Increasing the weight of LV from 1 to 2 led to noticeable gains in accuracy, F1 score, AUROC, and AUPRC."

    The weight λV=2 is not derived from first principles; it is selected because the sensitivity analysis on exactly the three target datasets that later appear in Table 2 shows that this value maximizes AUROC/AUPRC. The abstract claim that the method consistently outperforms other seven baselines is then reported for that selected configuration, so the headline comparison is not an independent test of the method: the reported superiority is partly constructed by choosing the hyperparameter on the same test metrics used to evaluate the baselines.

full rationale

The paper does not contain a mathematical derivation chain: DAAC is an empirical pipeline (Discrepancy Estimator + Adaptive Contrastive Learner + fine-tuning). No equation is shown to be equivalent to another input by construction. The DE reconstruction error E is a standard reconstruction-based anomaly score trained only on external normal data; calling it disease probability is an interpretive overclaim, but not circular. F.2's AUROC=0.974 for E on AD is an empirical validation, not a re-derivation. The ACL losses are standard InfoNCE/contrastive forms, and the COMET hierarchy is cited external work [8], not self-citation. No load-bearing self-citations appear. The main circularity-adjacent concern is empirical selection of the loss weights. Appendix D.2 describes a sensitivity analysis on the same three target datasets (AD, PTB, TDBrain) under the same 100% and 10% regimes used for the headline results, and Section 4.2 fixes λV=2 because that value improves the very metrics later used to claim superiority. This is a fitted hyperparameter presented as a general finding, and it makes the 'consistently outperforms' claim partially in-sample. However, the central method still has independent content: even with non-optimal weights (e.g., λV=1), ACL variants generally beat COMET in Table D1, so the result does not reduce entirely to the fit. I therefore set the circularity score at 4 rather than 6. Separately, the manuscript has internal inconsistencies that are correctness/reproducibility issues, not circularity: Table D1 and Table 2/E2 report different AUROC/AUPRC for identical configurations (e.g., AD 100% config (1,1,1,1,0): 93.60 vs 94.44; config (1,1,1,1,2): 96.40 vs 96.22), and Table 1 contradicts Section 5.1's 'excels in all metrics' because COMET+DE has higher AUROC/AUPRC than DAAC. These should be corrected, but they do not change the circularity assessment.

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

The headline numbers rest on one hand-tuned loss-weight configuration selected on the target datasets, an unvalidated-on-external-benchmarks discrepancy feature, and a set of domain-transfer assumptions that the paper itself shows can fail. No new physical or formal entities are postulated; the invented quantities are the discrepancy feature and the learned attention views, both evidenced only in-paper.

free parameters (5)
  • Contrastive loss weights lambda_S:lambda_R:lambda_E:lambda_T:lambda_V = 1:1:1:1:2
    Hand-chosen (Sec. 4.2) and justified post hoc by sensitivity analysis on the three target datasets (Appendix D.2, Table D1). The headline DAAC numbers use this best-on-target configuration; Sec. 6 says joint optimization is future work.
  • Contrastive temperature tau = not reported
    Appears in Eqs. 4, 5, 10 but no value is given anywhere in the paper.
  • Number of attention views V = 2
    Assumed in Sec. 4.2.5 ('Assume we have two views... V = 2'); no sensitivity analysis over V.
  • Random continuous masking ratio/length = not reported
    The epoch- and temporal-wise losses (Secs. 4.2.3, 4.2.4) depend on this augmentation, but its parameters are never specified.
  • Discrepancy feature design (sequence-level scalar MSE) = scalar MSE over time and channels
    Selected by comparing four options on the target datasets (Table F1), i.e., chosen on the same data used for headline results.
assumptions (5)
  • domain assumption Reconstruction error from a normal-data autoencoder is a valid proxy for disease probability in target data
    Sec. 4.1.3 (Eq. 2): 'the magnitude of E indirectly reflects the abnormality of the sample... the probability of disease.' Sec. C3 documents a failure case (AUROC drop 94.44 to 93.97 under domain shift).
  • domain assumption External normal datasets transfer to target datasets despite demographic/protocol differences
    Sec. 4.1 relies on D' and D sharing 'similar characteristics'; Sec. C3 concedes differences (mean ages 63.6 vs 72.5 across the two AD datasets).
  • domain assumption MHA heads, pushed apart by contrastive loss, yield complementary views
    Sec. 4.2.5 claims inter-view loss 'promoting the diversity of views'; only UMAP visualization is offered, no per-head analysis.
  • domain assumption Subject/trial IDs define valid positive pairs in LS and LR
    Eqs. 4-5. For TDBrain (Sec. C2.2) 'trials' are synthetic pseudo-trials from windowing, so the trial identity is an artificial construct.
  • standard math InfoNCE losses with these pair definitions produce useful representations
    Eqs. 4-10 follow [39]; standard contrastive machinery, not derived here.
invented entities (2)
  • Discrepancy feature E (reconstruction error as 'disease probability')
    purpose: Augments every target sample with one scalar (Eq. 3) to inject knowledge from external normal recordings into all later stages.
    Validity shown only in-paper: MI on the AD target set (Table E1) and anomaly AUROC 0.974 on the same 22-subject AD set (Sec. F.2), single dataset, no external benchmark.
  • Learned attention views (V=2)
    purpose: Discriminative 'views' used by inter/intra-view losses (Eqs. 9-10) to replace handcrafted contrastive pairs.
    Emergent constructs evidenced only by in-paper ablation (Table E2) and UMAP (Fig. 3); no external handle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Discrepancy-Aware Contrastive Adaptation in Medical Time Series Analysis." pith.science (2026). https://pith.science/paper/2P4YIXWA

@misc{pith2026250805572,
  author       = {Pith},
  title        = {Pith review of: Discrepancy-Aware Contrastive Adaptation in Medical Time Series Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2P4YIXWA}},
  note         = {Machine review of arXiv:2508.05572}
}
read the original abstract

In medical time series disease diagnosis, two key challenges are identified. First, the high annotation cost of medical data leads to overfitting in models trained on label-limited, single-center datasets. To address this, we propose incorporating external data from related tasks and leveraging AE-GAN to extract prior knowledge, providing valuable references for downstream tasks. Second, many existing studies employ contrastive learning to derive more generalized medical sequence representations for diagnostic tasks, usually relying on manually designed diverse positive and negative sample pairs. However, these approaches are complex, lack generalizability, and fail to adaptively capture disease-specific features across different conditions. To overcome this, we introduce LMCF (Learnable Multi-views Contrastive Framework), a framework that integrates a multi-head attention mechanism and adaptively learns representations from different views through inter-view and intra-view contrastive learning strategies. Additionally, the pre-trained AE-GAN is used to reconstruct discrepancies in the target data as disease probabilities, which are then integrated into the contrastive learning process. Experiments on three target datasets demonstrate that our method consistently outperforms other seven baselines, highlighting its significant impact on healthcare applications such as the diagnosis of myocardial infarction, Alzheimer's disease, and Parkinson's disease. We release the source code at xxxxx.

Figures

Figures reproduced from arXiv: 2508.05572 by the authors.

Figure 1
Figure 1. Overview of our three-stage training framework. We first pretrain a Discrepancy Estimator ( [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Mechanism of Multi-View Contrastive Learning Furthermore, to capture the intrinsic hierarchical information embedded in medical data, we adopt the concept from COMET [8]. Our contrastive learning considers subject, trial, epoch, and temporal, and optimizes the model using multiple InfoNCE losses [39]. Ultimately, we obtain representations that are beneficial for downstream medical time-series tasks. 4.2.1 Subject-wi… view at source ↗
Figure 3
Figure 3. Visualization of Multi-View Representations on the AD Dataset. The figure compares contrastive view [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 39 canonical work pages

  1. [1]

    Deep state space models for time series forecasting

    Syama Sundar Rangapuram, Matthias W Seeger, Jan Gasthaus, Lorenzo Stella, Yuyang Wang, and Tim Januschowski. Deep state space models for time series forecasting. Advances in neural information processing systems, 31, 2018. 8 Discrepancy-Aware Adaptive Contrastive learnin

    work page 2018

  2. [2]

    Auto-regressive moving diffusion models for time series forecasting

    Jiaxin Gao, Qinglong Cao, and Yuntian Chen. Auto-regressive moving diffusion models for time series forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 16727–16735, 2025

    work page 2025

  3. [3]

    Kandinsky conformal prediction: Efficient calibration of image segmentation algorithms

    Joren Brunekreef, Eric Marcus, Ray Sheombarsing, Jan-Jakob Sonke, and Jonas Teuwen. Kandinsky conformal prediction: Efficient calibration of image segmentation algorithms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4135–4143, 2024

    work page 2024

  4. [4]

    Time series data augmentation for deep learning: A survey

    Qingsong Wen, Liang Sun, Fan Yang, Xiaomin Song, Jingkun Gao, Xue Wang, and Huan Xu. Time series data augmentation for deep learning: A survey. arXiv preprint arXiv:2002.12478, 2020

    arXiv 2002

  5. [5]

    Deep learning for time series forecasting: a survey

    José F Torres, Dalil Hadjout, Abderrazak Sebaa, Francisco Martínez-Álvarez, and Alicia Troncoso. Deep learning for time series forecasting: a survey. Big data, 9(1):3–21, 2021

    work page 2021

  6. [6]

    Self-supervised learning: Generative or contrastive

    Xiao Liu, Fanjin Zhang, Zhenyu Hou, Li Mian, Zhaoyu Wang, Jing Zhang, and Jie Tang. Self-supervised learning: Generative or contrastive. IEEE transactions on knowledge and data engineering, 35(1):857–876, 2021

    work page 2021

  7. [7]

    Mp-net: a multi-center privacy-preserving network for medical image segmentation

    Enjun Zhu, Haiyu Feng, Long Chen, Yongqiang Lai, and Senchun Chai. Mp-net: a multi-center privacy-preserving network for medical image segmentation. IEEE Transactions on Medical Imaging, 2024

    work page 2024

  8. [8]

    Contrast everything: A hierarchical contrastive framework for medical time-series

    Yihe Wang, Yu Han, Haishuai Wang, and Xiang Zhang. Contrast everything: A hierarchical contrastive framework for medical time-series. Advances in Neural Information Processing Systems, 36, 2024

    work page 2024

Show all 48 references
  1. [9]

    Clocs: Contrastive learning of cardiac signals across space, time, and patients

    Dani Kiyasseh, Tingting Zhu, and David A Clifton. Clocs: Contrastive learning of cardiac signals across space, time, and patients. In International Conference on Machine Learning, pages 5606–5615. PMLR, 2021

    2021

  2. [10]

    Self-supervised contrastive learning for medical time series: A systematic review

    Ziyu Liu, Azadeh Alavi, Minyi Li, and Xiang Zhang. Self-supervised contrastive learning for medical time series: A systematic review. Sensors, 23(9):4221, 2023

    2023

  3. [11]

    Unbiased classification through bias-contrastive and bias-balanced learning

    Youngkyu Hong and Eunho Yang. Unbiased classification through bias-contrastive and bias-balanced learning. Advances in Neural Information Processing Systems, 34:26449–26461, 2021

    2021

  4. [12]

    Selecmix: Debiased learning by contradicting-pair sampling

    Inwoo Hwang, Sangjun Lee, Yunhyeok Kwak, Seong Joon Oh, Damien Teney, Jin-Hwa Kim, and Byoung-Tak Zhang. Selecmix: Debiased learning by contradicting-pair sampling. Advances in Neural Information Processing Systems, 35:14345–14357, 2022

    2022

  5. [13]

    Federated active learning for multicenter collaborative disease diagnosis

    Xing Wu, Jie Pei, Cheng Chen, Yimin Zhu, Jianjia Wang, Quan Qian, Jian Zhang, Qun Sun, and Yike Guo. Federated active learning for multicenter collaborative disease diagnosis. IEEE transactions on medical imaging, 42(7):2068–2080, 2022

  6. [14]

    A 30-min nucleic acid amplification point-of-care test for genital chlamydia trachomatis infection in women: a prospective, multi-center study of diagnostic accuracy

    Emma M Harding-Esch, Emma C Cousins, S-LC Chow, Laura T Phillips, Catherine L Hall, Nicholas Cooper, Sebastian S Fuller, Achyuta V Nori, Rajul Patel, Sathish Thomas-William, et al. A 30-min nucleic acid amplification point-of-care test for genital chlamydia trachomatis infecti...

    2018

  7. [15]

    The mechanism of auditory evoked eeg responses

    B Mca Savers, HA Beagley, and WR Henshall. The mechanism of auditory evoked eeg responses. Nature, 247(5441):481–483, 1974

    1974

  8. [16]

    Ai-enabled electrocardiography alert intervention and all-cause mortality: a pragmatic randomized clinical trial

    Chin-Sheng Lin, Wei-Ting Liu, Dung-Jang Tsai, Yu-Sheng Lou, Chiao-Hsiang Chang, Chiao-Chin Lee, Wen-Hui Fang, Chih-Chia Wang, Yen-Yuan Chen, Wei-Shiang Lin, et al. Ai-enabled electrocardiography alert intervention and all-cause mortality: a pragmatic randomized clinical trial....

    2024

  9. [17]

    A survey on hand gesture recognition based on surface electromyography: Fundamentals, methods, applications, challenges and future trends

    Sike Ni, Mohammed AA Al-qaness, Ammar Hawbani, Dalal Al-Alimi, Mohamed Abd Elaziz, and Ahmed A Ewees. A survey on hand gesture recognition based on surface electromyography: Fundamentals, methods, applications, challenges and future trends. Applied Soft Computing, page 112235, 2024

    2024

  10. [18]

    Design and development of human computer interface using electrooculogram with deep learning

    Geer Teng, Yue He, Hengjun Zhao, Dunhu Liu, Jin Xiao, and S Ramkumar. Design and development of human computer interface using electrooculogram with deep learning. Artificial intelligence in medicine, 102:101765, 2020

    2020

  11. [19]

    Ts2vec: Towards universal representation of time series

    Zhihan Yue, Yujing Wang, Juanyong Duan, Tianmeng Yang, Congrui Huang, Yunhai Tong, and Bixiong Xu. Ts2vec: Towards universal representation of time series. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 8980–8987, 2022

    2022

  12. [20]

    Anomaly transformer: Time series anomaly detection with association discrepancy

    Guoqiang Xu, Weize Zhang, Yujing Zhang, Ziqing Xiong, Zhengyu Deng, and Lejian Sun. Anomaly transformer: Time series anomaly detection with association discrepancy. In International Conference on Learning Representations (ICLR), 2022

    2022

  13. [21]

    Dabaclt: A data augmentation bias-aware contrastive learning framework for time series representation

    Yu Zheng, Haokun Lin, Zicheng Li, Zhaowei Wang, and Yiming Zhao. Dabaclt: A data augmentation bias-aware contrastive learning framework for time series representation. Applied Sciences, 13(13):7908, 2023

    2023

  14. [22]

    Time series contrastive learning with information-aware augmenta- tions

    Chenghao Luo, Ziyue Wang, Hao Zhang, and Liang Zhao. Time series contrastive learning with information-aware augmenta- tions. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 9744–9752, 2023

    2023

  15. [23]

    Discrepancy-based evolutionary diversity optimization

    Aneta Neumann, Wanru Gao, Carola Doerr, Frank Neumann, and Markus Wagner. Discrepancy-based evolutionary diversity optimization. In Proceedings of the Genetic and Evolutionary Computation Conference, pages 991–998, 2018

    2018

  16. [24]

    U-sleep: resilient high-frequency sleep staging

    Mathias Perslev, Sune Darkner, Lykke Kempfner, Miki Nikolic, Poul Jørgen Jennum, and Christian Igel. U-sleep: resilient high-frequency sleep staging. NPJ digital medicine, 4(1):72, 2021

    2021

  17. [25]

    L-seqsleepnet: Whole-cycle long sequence modeling for automatic sleep staging

    Huy Phan, Kristian P Lorenzen, Elisabeth Heremans, Oliver Y Chén, Minh C Tran, Philipp Koch, Alfred Mertins, Mathias Baumert, Kaare B Mikkelsen, and Maarten De V os. L-seqsleepnet: Whole-cycle long sequence modeling for automatic sleep staging. IEEE Journal of Biomedical and H...

    2023

  18. [26]

    Multi-scale inception-based deep fusion network for electrooculogram-based eye movements classification

    Zheng Zeng, Linkai Tao, Jun Hu, Ruizhi Su, Long Meng, Chen Chen, and Wei Chen. Multi-scale inception-based deep fusion network for electrooculogram-based eye movements classification. Biomedical Signal Processing and Control, 103:107377, 2025

    2025

  19. [27]

    Simgans: Simulator-based generative adversarial networks for ecg synthesis to improve deep ecg classification

    Tomer Golany, Kira Radinsky, and Daniel Freedman. Simgans: Simulator-based generative adversarial networks for ecg synthesis to improve deep ecg classification. In International Conference on Machine Learning, pages 3597–3606. PMLR, 2020

    2020

  20. [28]

    Generative adversarial networks in electrocardiogram synthesis: Recent developments and challenges

    Laurenz Berger, Max Haberbusch, and Francesco Moscato. Generative adversarial networks in electrocardiogram synthesis: Recent developments and challenges. Artificial Intelligence in Medicine, 143:102632, 2023

    2023

  21. [29]

    Imbalanced domain adaptation net for multi-class electrocardiography signals classification

    Ran An, Zuogang Shang, Yuhong Liu, Xin Wu, Jing Ren, and Zhibin Zhao. Imbalanced domain adaptation net for multi-class electrocardiography signals classification. Biomedical Signal Processing and Control, 108:107912, 2025

    2025

  22. [30]

    Cross-domain industrial intrusion detection deep model trained with imbalanced data

    Yongle Chen, Sida Su, Dan Yu, Hao He, Xiaojian Wang, Yao Ma, and Hao Guo. Cross-domain industrial intrusion detection deep model trained with imbalanced data. IEEE Internet of Things Journal, 10(1):584–596, 2022

    2022

  23. [31]

    Self-supervised contrastive pre-training for time series via time-frequency consistency

    Xiang Zhang, Ziyuan Zhao, Theodoros Tsiligkaridis, and Marinka Zitnik. Self-supervised contrastive pre-training for time series via time-frequency consistency. Advances in Neural Information Processing Systems, 35:3988–4003, 2022

    2022

  24. [32]

    Robust time series recovery and classification using test-time noise simulator networks

    Eun Som Jeon, Suhas Lohit, Rushil Anirudh, and Pavan Turaga. Robust time series recovery and classification using test-time noise simulator networks. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2023

    2023

  25. [33]

    Addressing spatial-temporal heterogeneity: General mixed time series analysis via latent continuity recovery and alignment

    Jiawei Chen. Addressing spatial-temporal heterogeneity: General mixed time series analysis via latent continuity recovery and alignment. Advances in Neural Information Processing Systems, 37:17910–17946, 2024

    2024

  26. [34]

    Self-supervised learning for time series analysis: Taxonomy, progress, and prospects

    Kexin Zhang, Qingsong Wen, Chaoli Zhang, Rongyao Cai, Ming Jin, Yong Liu, James Y Zhang, Yuxuan Liang, Guansong Pang, Dongjin Song, et al. Self-supervised learning for time series analysis: Taxonomy, progress, and prospects. IEEE Transactions on Pattern Analysis and Machine In...

    2024

  27. [35]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597–1607. PMLR, 2020

    2020

  28. [36]

    Self- supervised contrastive representation learning for semi-supervised time-series classification

    Emadeldeen Eldele, Mohamed Ragab, Zhenghua Chen, Min Wu, Chee-Keong Kwoh, Xiaoli Li, and Cuntai Guan. Self- supervised contrastive representation learning for semi-supervised time-series classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023

    2023

  29. [37]

    Timesurl: Self-supervised contrastive learning for universal time series representation learning

    Jiexi Liu and Songcan Chen. Timesurl: Self-supervised contrastive learning for universal time series representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 13918–13926, 2024

    2024

  30. [38]

    Soft contrastive learning for time series

    Seunghan Lee, Taeyoung Park, and Kibok Lee. Soft contrastive learning for time series. arXiv preprint arXiv:2312.16424, 2023

    2023

  31. [39]

    Representation learning with contrastive predictive coding

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018

    2018 arXiv

  32. [40]

    Time-series representation learning via temporal and contextual contrasting

    Emadeldeen Eldele, Mohamed Ragab, Zhenghua Chen, Min Wu, Chee Keong Kwoh, Xiaoli Li, and Cuntai Guan. Time-series representation learning via temporal and contextual contrasting. arXiv preprint arXiv:2106.14112, 2021

    2021 arXiv

  33. [41]

    Detecting anomalies within time series using local neural transformations

    Tim Schneider, Chen Qiu, Marius Kloft, Decky Aspandi Latif, Steffen Staab, Stephan Mandt, and Maja Rudolph. Detecting anomalies within time series using local neural transformations. arXiv preprint arXiv:2202.03944, 2022

    2022 arXiv

  34. [42]

    Umap: Uniform manifold approximation and projection for dimension reduction

    Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018

    2018 arXiv

  35. [43]

    A dataset of eeg recordings from: Alzheimer’s disease, frontotemporal dementia and healthy subjects

    Andreas Miltiadous, Katerina D Tzimourta, Theodora Afrantou, Panagiotis Ioannidis, Nikolaos Grigoriadis, Dimitrios G Tsalikakis, Pantelis Angelidis, Markos G Tsipouras, Evripidis Glavas, Nikolaos Giannakeas, et al. A dataset of eeg recordings from: Alzheimer’s disease, frontot...

    2023

  36. [44]

    Linear predictive coding distinguishes spectral eeg features of parkinson’s disease.Parkinsonism & related disorders, 79:79–85, 2020

    Md Fahim Anjum, Soura Dasgupta, Raghuraman Mudumbai, Arun Singh, James F Cavanagh, and Nandakumar S Narayanan. Linear predictive coding distinguishes spectral eeg features of parkinson’s disease.Parkinsonism & related disorders, 79:79–85, 2020

    2020

  37. [45]

    Ptb-xl, a large publicly available electrocardiography dataset

    Patrick Wagner, Nils Strodthoff, Ralf-Dieter Bousseljot, Dieter Kreiseler, Fatima I Lunze, Wojciech Samek, and Tobias Schaeffter. Ptb-xl, a large publicly available electrocardiography dataset. Scientific data, 7(1):1–15, 2020

    2020

  38. [46]

    Analysis of electroencephalograms in alzheimer’s disease patients with multiscale entropy

    J Escudero, Daniel Abásolo, Roberto Hornero, Pedro Espino, and Miguel López. Analysis of electroencephalograms in alzheimer’s disease patients with multiscale entropy. Physiological measurement, 27(11):1091, 2006

    2006

  39. [47]

    The two decades brainclinics research archive for insights in neurophysiology (tdbrain) database

    Hanneke Van Dijk, Guido Van Wingen, Damiaan Denys, Sebastian Olbrich, Rosalinde Van Ruth, and Martijn Arns. The two decades brainclinics research archive for insights in neurophysiology (tdbrain) database. Scientific data, 9(1):333, 2022

    2022

  40. [48]

    Nutzung der ekg-signaldatenbank cardiodat der ptb über das internet

    Ralf-Dieter Bousseljot, Dieter Kreiseler, and Andreas Schnabel. Nutzung der ekg-signaldatenbank cardiodat der ptb über das internet. Biomedizinische Technik/Biomedical Engineering, 40(Supplement 1):317–318, 1995. 10 Discrepancy-Aware Adaptive Contrastive learnin A Preliminary ...

    1995

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.