Pith. sign in

REVIEW 4 major objections 5 minor 32 references

Multi-scale Masked Autoencoder for Electrocardiogram Anomaly Detection

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

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

Pith's one-line read A lightweight multi-scale masked autoencoder matches state-of-the-art ECG anomaly detection without needing R-peak detection or heartbeat segmentation.

desk verdict A practical-looking R-peak-free ECG anomaly detector whose main claim is undercut by an unmasked copying path in the masking scheme. read the letter →

arxiv 2502.05494 v1 pith:AALJX2PW submitted 2025-02-08 cs.LG cs.AIstat.AP

Ya Zhou , Yujie Yang , Jianhuang Gan , Xiangjie Li , Jing Yuan , Wei Zhao This is my paper · ORCID
classification cs.LGcs.AIstat.AP
keywords electrocardiogramanomalydetectionmaskedautoencodermulti-scaleattentionTransformerself-supervisedlearningPTB-XLlocalizationreconstructionerror
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

This paper claims that ECG anomaly detection can be done end-to-end by a multi-scale masked autoencoder trained on normal recordings only, without R-peak detection or heartbeat segmentation. The method splits a 12-lead ECG into non-overlapping segments, masks both a global set of segments and a local region in each training pass, and reconstructs the masked segments with a lightweight Transformer. On the PTB-XL benchmark it reports detection AUC 0.860 and localization AUC 0.749, matching the previous state of the art while using 0.576 GFLOPs per inference versus 45.108 GFLOPs and 0.398M parameters versus 7.086M. The payoff is a preprocessing-free, low-compute pipeline that could make ECG anomaly screening more practical in clinical settings.

What carries the argument

The load-bearing mechanism is multi-scale masking combined with multi-scale cross-attention and distinct positional embeddings. Masking is applied separately to the full segment sequence (global scale) and to a randomly chosen local window of consecutive segments (local scale), so the model is forced to reconstruct both whole-signal morphology and short, beat-like details; distinct positional embeddings tell the encoder whether a token came from the global or local scale. The decoder is a single Transformer block that reconstructs per-segment normalized values, and inference repeats several random forward passes, aggregating losses over all local regions so every point is masked with high probability. The anomaly score is the mean squared reconstruction error of masked segments, and localization is the same score attributed to individual signal points.

What would settle it

Train the identical model on normal recordings, then measure the distribution of the anomaly score on held-out normal and abnormal ECGs; if the overlapping area of the two score distributions implies an AUC at or below 0.5 on a new cohort, the central claim collapses. A sharper test is to inject synthetic noise or an unseen non-cardiac artifact into normal signals: if reconstruction error for those normal-but-noisy samples rises above the abnormal threshold, the score is capturing novelty rather than cardiac pathology.

Watch

Extended reading notes

Core claim

The central discovery is that masked autoencoding, previously reported to underperform at unsupervised image anomaly detection, can match state-of-the-art ECG anomaly detection when masking and attention are made multi-scale. The paper partitions each signal into non-overlapping segments, assigns separate learnable positional embeddings for global and local representations, and during training randomly masks both a global set of segments and segments inside a local window of consecutive segments. A three-layer Transformer encoder cross-attends over the unmasked global and local segments, and a single-layer decoder reconstructs the normalized masked segments; the per-sample anomaly score is the average reconstruction loss over all local regions and forward passes, and per-point scores are obtained by aggregating the same losses over signal points. With this design the paper reports detection AUC 0.860 and localization AUC 0.749 on PTB-XL, matching the prior state-of-the-art detection score and slightly surpassing its localization score, while removing the R-peak preprocessing dependency that earlier methods require.

Load-bearing premise

The whole pipeline rests on the premise that, after multi-scale masked training on normal ECGs, an abnormal waveform is harder to reconstruct than a normal one, and that this difference survives the per-segment normalization used to compute the reconstruction loss.

Editorial extensions

If this is right

  • If the reported numbers hold, ECG anomaly screening no longer needs a preprocessing stage for R-peak detection or heartbeat segmentation, removing a common failure point on noisy or pathological recordings.
  • The 78x reduction in inference FLOPs and the 0.398M parameter count make the method deployable on resource-constrained hardware such as bedside monitors or portable ECG devices.
  • The result challenges the earlier finding that masked autoencoders are poorly suited to unsupervised anomaly detection, at least for structured physiological signals.
  • Because the model trains on normal data only, it preserves the ability to flag rare or unseen abnormality types rather than only those in a labeled training set.
  • The per-point reconstruction loss gives a localization map alongside the sample-level decision, so the same forward pass supports both detection and interpretation.

Reading between the lines

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

  • The same multi-scale masking recipe could transfer to other quasi-periodic physiological signals, such as fetal heart rate or arterial blood pressure waveforms, where beat segmentation is equally fragile.
  • An implicit testable extension is whether the per-segment normalization and the aggregation over forward passes can be tuned per lead; the paper does not analyze lead-wise score calibration.
  • A direct stress test would be to run the same pipeline on noisy or ambulatory recordings; if reconstruction-error separation degrades, the method's clinical robustness claim would need qualification.
  • The ablation results suggest the global and local scales contribute complementary information, which could motivate a principled weighting scheme instead of the equal averaging used in the anomaly score formula.
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 manuscript proposes MMAE-ECG, a multi-scale masked autoencoder for ECG anomaly detection and localization. The method partitions a 12-lead ECG into non-overlapping global segments, defines overlapping local regions, masks both scales independently during training, and reconstructs the masked segments with a lightweight Transformer encoder-decoder. At inference, the anomaly score at both sample and point level is the averaged reconstruction loss over multiple forward passes and local regions. On the PTB-XL benchmark, the authors report a detection AUC of 0.860 and a localization AUC of 0.749, matching MCF (0.860/0.747) while claiming a roughly 78x reduction in inference FLOPs (0.576 vs 45.108 GFLOPs) and a much smaller model (0.398M vs 7.086M parameters). Ablation studies show degradation when the global region, local region, local positional embedding, multi-scale masking, or masked-segment-only loss are removed.

Significance. If the reported results are valid, MMAE-ECG would be a meaningful contribution: it removes the R-peak detection and heartbeat-segmentation preprocessing required by prior ECG anomaly detectors, is substantially lighter than MCF, and achieves competitive detection and localization on a public benchmark. The paper's strengths are its use of the public PTB-XL benchmark, a clear ablation study covering several architectural choices, and a concrete efficiency comparison. However, the evidence as presented is not yet sufficient: the central comparison rests on a single table without error bars or multiple seeds, no code is provided, and the masking formulation as written permits a trivial self-copying path in the decoder that can undermine the intended anomaly score. The announced efficiency gain also relies on a FLOPs comparison that is not obviously apples-to-apples.

major comments (4)
  1. [Sections 3.1 and 3.3, Eq. (1)] The multi-scale masking as defined permits the decoder to copy a masked segment's own content rather than predicting it from context. The global mask U_mask and the local mask V^w_mask are sampled independently, and the text only enforces disjointness within each scale (U_unmask ∩ U_mask = ∅ and V^w_unmask ∩ V^w_mask = ∅), not between scales. Since V^w ⊂ U, with T=40, θ=0.25, S=10, a segment that is local-masked is globally unmasked with probability 1 − S/T = 0.75. That segment's encoded representation z_L^{i_s} is then placed in the decoder input alongside its own masked token (Section 3.3), so the decoder's self-attention can copy the segment's content. The same holds symmetrically for a global-masked segment that is in V^w_unmask. The paper never states that overlapping segments are excluded from the loss or that U_mask ∩ V^w_mask = ∅. Because Eq. (1) is the entire anomaly score and the local loss drives the 0.749 localization AUC, the reported numbers may reflect copy fidelity rather than the intended multi-scale reconstruction difficulty. Please fix the masking so that no segment is simultaneously visible and masked in the same forward pass, or explicitly remove overlapping visible tokens from the decoder input, and rerun the experiments.
  2. [Table 1 and Section 4.2] The central comparison to state-of-the-art methods is reported from a single run without error bars, confidence intervals, or multiple seeds. The detection AUC is 0.860 for both MCF and MMAE-ECG, and the localization difference is 0.749 vs 0.747; with no uncertainty estimate, phrases such as 'comparable' and 'slightly better' are not supported. Please report the mean and standard deviation over at least five independent training runs, include per-run values or a significance test, and release the code needed to reproduce Table 1 and Table 3.
  3. [Section 4.2, Table 2] The FLOPs comparison is not obviously measuring the same workload. MCF's 45.108 GFLOPs is computed as 1.253 × 12 × 3, where 1.253 GFLOPs is per forward pass, 12 is the median number of R-peaks, and 3 accounts for feed-forward operations. MMAE-ECG's 0.576 GFLOPs is 0.016 × 9 × 4, using the number of local regions and H = 4. It is unclear whether the per-pass figure of 0.016 includes the full encoder-decoder computation for all global and local tokens, and whether MCF's per-pass overhead is measured on the same input length and hardware. Since the 78x efficiency claim is a major advertised contribution, please provide a standardized FLOPs measurement, such as profiling both models on the same 10-second, 12-lead input with a public FLOPs counter, and state the exact tensor shapes and operations included.
  4. [Section 3.3] The reconstruction loss uses a 'predefined per-segment normalization function f as specified in Zhou et al. (2023)', but the paper does not define f. This is not merely a reproducibility issue: if f is a per-segment normalization that rescales each segment independently, it may remove amplitude information that is critical for detecting morphological anomalies. The authors should state the exact form of f, whether it is trainable or fixed, and discuss how it preserves or discards the abnormality signal that the anomaly score is supposed to capture.
minor comments (5)
  1. [Section 3.1] The symbol S is used with two conflicting meanings: it is first defined as the number of masked global segments (S := min{max{[Tθ],1}, T−1}), but later used as the index upper bound for the unmasked set U_unmask = {X_{i_1},...,X_{i_S}}. With θ=0.25 and T=40, S=10 masked segments would leave 30 unmasked, not 10. Please rename one of the two counts.
  2. [Section 3.4] The probability that a segment is masked at least once in H passes is 1 − (1 − R/δ)^H, not 1 − (R/δ)^H as written. Please correct the formula.
  3. [Section 4.4 and Figure 3A] The text says the masking ratio experiments range 'from 0.5 to 0.95', but the horizontal axis in Figure 3A starts at 0.05 and the reported AUC values correspond to ratios below 0.5. This appears to be a typo; please fix the range description.
  4. [Abstract and Section 5] The abstract says performance is 'comparable to state-of-the-art approaches', while the Discussion says the method 'achieves state-of-the-art performance' and 'superior performance'. Please choose one characterization and use it consistently, since the table shows equal detection AUC and a marginal localization difference.
  5. [Section 5] The sentence 'To the best of our knowledge, this is the first work to integrate these techniques' is an overclaim given the proximity to Zhou et al. (2023) and Jiang et al. (2023); please soften it or provide a more precise novelty statement.

Circularity Check

1 steps flagged · score 6.0 of 10

Cross-scale mask overlap lets local-masked segments feed their own encodings into the decoder, so Eq. (1)'s local loss partly measures copying rather than anomaly.

  1. other [Section 3.1 (Multi-scale Masking) and Section 3.3 (Multi-Scale Reconstruction)]
    "we randomly select w ∈ {w1, . . . , wν} and separately apply masking to the elements in U and V w ... U = Uunmask ∪ Umask, with Uunmask ∩ Umask = ∅, and V w = V w unmask ∪ V w mask, with V w unmask ∩ V w mask = ∅. Uunmask and V w unmask are fed into the encoder to achieve multi-scale cross-attention, while Umask and V w mask serve as the reconstruction targets."

    By the paper's own definitions, U := {X1,...,XT} and V^w := {X_{w+1},...,X_{w+δ}}, so V^w is a subset of U. Disjointness is enforced only inside each scale, not between U_unmask and V^w_mask. With θ=0.25, T=40, δ=4, R=1, a local-masked segment is not chosen for global masking with probability 1 − 10/40 = 0.75, so it belongs to U_unmask and is encoded. In Section 3.3 the decoder input concatenates that encoded token z_L^{i_s} together with the masked token e_m for the same segment. The one-layer Transformer decoder can then attend to the segment's own encoded content and copy it, making l^w_local in Eq. (1) a reconstruction of information already given to the model rather than a prediction from context.

full rationale

This is otherwise an empirical benchmark paper: the anomaly score is the mean reconstruction loss, the reported AUCs come from the external PTB-XL benchmark, no fitted parameter is relabeled as a prediction, and the self-citations to Zhou et al. (2023) for segmentation, per-segment normalization, and positional embeddings are implementation choices rather than unverified uniqueness theorems. However, the masking construction contains a genuine internal circularity: because local windows are subsets of the global segmentation and masking is applied independently at the two scales, most local-masked segments remain unmasked globally and their own encoded representations are placed in the decoder input. For those segments the reconstruction target is also an input, so Eq. (1)'s local loss is partly a copy task. This does not make the entire comparison vacuous—global reconstruction and the non-leaked local cases still carry information—but it means a substantial component of the reported anomaly detection and localization signal is circular by the paper's own equations.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

No new physical or conceptual entities are introduced; the framework is a composition of known components (MAE, multi-scale attention, positional embeddings). The load-bearing postulates are the reconstruction-error-as-anomaly-score premise, the adequacy of normal-only training, the correctness of the external benchmark, and the chosen normalization and Monte Carlo aggregation.

free parameters (6)
  • Segment count T and segment size Q/T = T=40, segment size 125 samples at 500 Hz
    Partitions each 10 s, 12-lead ECG into non-overlapping segments; determines the granularity of global and local features and is carried over from Zhou et al. (2023).
  • Local region length delta = delta=4 segments
    Number of consecutive segments forming a local region; hand chosen to capture a few heartbeats worth of local morphology.
  • Local region anchor positions = {1, 5, 9, 13, 17, 21, 25, 29, 33}
    Starting indices for the 9 local regions; 'similar to Jiang et al. (2023)' (Section 4.1), i.e., chosen by the benchmark design.
  • Masking ratio theta = 0.25
    Fraction of segments masked; ablation in Figure 3A shows 0.15-0.35 is optimal, so 0.25 is selected on the benchmark.
  • Inference aggregation passes H = 4
    Number of random forward passes at inference; chosen so each local segment is masked with at least 99% probability, and Figure 3B shows AUC varies with H.
  • Encoder and decoder dimensions = L=3 layers, 16 heads, D=64; decoder D'=64, 2 heads
    Model capacity hyperparameters; no sensitivity analysis is reported for these.
assumptions (5)
  • domain assumption Reconstruction error on masked ECG segments separates normal from abnormal signals (Eq. 1).
    The anomaly score is the average reconstruction MSE over global and local losses; if abnormal signals were not harder to reconstruct, the AUC claim would fail.
  • domain assumption Normal-only training data (8,167 normal PTB-XL recordings) is sufficient to model normal ECG variability.
    Anomaly detection is trained exclusively on normal data, as stated in the introduction and Section 4.
  • domain assumption PTB-XL detection and localization labels and benchmark protocol of Jiang et al. (2023) are correct and directly comparable.
    All reported metrics are computed on this benchmark and comparisons with MCF and TSRNet are transcribed from those papers.
  • ad hoc to paper Per-segment normalization f adopted from Zhou et al. (2023) does not erase the anomaly information needed for localization.
    The reconstruction target is normalized segments (Section 3.3); if f normalizes away abnormal waveform morphology, point-level localization scores would lose signal.
  • ad hoc to paper Random masking of one local region per pass, averaged over H passes and 9 regions, yields stable point-level anomaly scores.
    Section 3.4 defines the score as a Monte Carlo average; the paper does not prove coverage or stability beyond the chosen H=4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-scale Masked Autoencoder for Electrocardiogram Anomaly Detection." pith.science (2026). https://pith.science/paper/AALJX2PW

@misc{pith2026250205494,
  author       = {Pith},
  title        = {Pith review of: Multi-scale Masked Autoencoder for Electrocardiogram Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AALJX2PW}},
  note         = {Machine review of arXiv:2502.05494}
}
read the original abstract

Electrocardiogram (ECG) analysis is a fundamental tool for diagnosing cardiovascular conditions, yet anomaly detection in ECG signals remains challenging due to their inherent complexity and variability. We propose Multi-scale Masked Autoencoder for ECG anomaly detection (MMAE-ECG), a novel end-to-end framework that effectively captures both global and local dependencies in ECG data. Unlike state-of-the-art methods that rely on heartbeat segmentation or R-peak detection, MMAE-ECG eliminates the need for such pre-processing steps, enhancing its suitability for clinical deployment. MMAE-ECG partitions ECG signals into non-overlapping segments, with each segment assigned learnable positional embeddings. A novel multi-scale masking strategy and multi-scale attention mechanism, along with distinct positional embeddings, enable a lightweight Transformer encoder to effectively capture both local and global dependencies. The masked segments are then reconstructed using a single-layer Transformer block, with an aggregation strategy employed during inference to refine the outputs. Experimental results demonstrate that our method achieves performance comparable to state-of-the-art approaches while significantly reducing computational complexity-approximately 1/78 of the floating-point operations (FLOPs) required for inference. Ablation studies further validate the effectiveness of each component, highlighting the potential of multi-scale masked autoencoders for anomaly detection.

Figures

Figures reproduced from arXiv: 2502.05494 by the authors.

Figure 1
Figure 1. Overview of the proposed framework. (1) Multi-scale Masking: Segments in the global and local regions are masked separately. (2) Multi-scale Cross-attention Encoding: Unmasked seg￾ments from both regions are concatenated and fed into a lightweight Transformer-based encoder for cross-attention. (3) Multi-scale Reconstruction: Masked segments in global and local regions are reconstructed using a single-layer Transform… view at source ↗
Figure 2
Figure 2. Examples of anomaly localization on PTB-XL with different abnormal types. Ground truths are highlighted in red boxes on the ECG signals, and anomaly localization of the proposed method are attached below. 4.4 Ablation Study We conduct ablation studies to systematically evaluate the contribution of each design choice in our model, using the PTB-XL anomaly detection benchmark, which includes patients with diverse char… view at source ↗
Figure 3
Figure 3. Ablation study results for different masking ratios and values of H. et al., 2024), our method eliminates the need for heartbeat segmentation or R-peak detection during the data preprocessing stage, which can be inconvenient and unreliable for certain ECG signals in real-world applications. Moreover, the proposed framework is lightweight, with only 0.398M parameters and 0.576 GFLOPs for inference, approximately 1/78… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 26 canonical work pages

  1. [1]

    L., Kiros, J

    Ba, J. L., Kiros, J. R. and Hinton, G. E. (2016) Layer normalization. arXiv preprint arXiv:1607.06450

    arXiv 2016

  2. [2]

    and Franklin, M

    Boniol, P., Paparrizos, J., Palpanas, T. and Franklin, M. J. (2021) Sand: streaming subsequence anomaly detection. Proc. VLDB Endow., 14, 1717–1729. ://doi.org/10.14778/3467861.3467863

    work page arXiv 2021

  3. [3]

    and Le, N

    Bui, N.-T., Hoang, D.-H., Phan, T., Tran, M.-T., Patel, B., Adjeroh, D. and Le, N. (2024) Tsrnet: Simple framework for real-time ecg anomaly detection with multimodal time and spectrogram restoration network. In 2024 IEEE International Symposium on Biomedical Imaging (ISBI), 1--4. IEEE

    work page 2024

  4. [4]

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S. et al. (2020) An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929

    arXiv 2020

  5. [5]

    and Girshick, R

    He, K., Chen, X., Xie, S., Li, Y., Doll \'a r, P. and Girshick, R. (2022) Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 16000--16009

    work page 2022

  6. [6]

    o derstr \

    Hundman, K., Constantinou, V., Laporte, C., Colwell, I. and S \"o derstr \"o m, T. (2018) Detecting spacecraft anomalies using lstms and nonparametric dynamic thresholding. Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. ://api.semanticscholar.org/CorpusID:3625298

    work page 2018

  7. [7]

    and Wang, Y

    Jiang, A., Huang, C., Cao, Q., Wu, S., Zeng, Z., Chen, K., Zhang, Y. and Wang, Y. (2023) Multi-scale cross-restoration framework for electrocardiogram anomaly detection. In International Conference on Medical Image Computing and Computer-Assisted Intervention, 87--97. Springer

    work page 2023

  8. [8]

    Self-supervised Anomaly Detection Pretraining Enhances Long-tail ECG Diagnosis

    Jiang, A., Huang, C., Cao, Q., Xu, Y., Zeng, Z., Chen, K., Zhang, Y. and Wang, Y. (2024) Self-supervised anomaly detection pretraining enhances long-tail ecg diagnosis. arXiv preprint arXiv:2408.17154

    work page Pith review arXiv 2024

Show all 32 references
  1. [9]

    and Cheng, X

    Liu, S., Zhou, B., Ding, Q., Hooi, B., Zhang, Z., Shen, H. and Cheng, X. (2022) Time series anomaly detection with adversarial reconstruction networks. IEEE Transactions on Knowledge and Data Engineering, 35, 4293--4306

    2022

  2. [10]

    C., Zhao, Z., Relton, S

    Qin, J., Gao, F., Wang, Z., Wong, D. C., Zhao, Z., Relton, S. D. and Fang, H. (2023) A novel temporal generative adversarial network for electrocardiography anomaly detection. Artificial Intelligence in Medicine, 136, 102489

    2023

  3. [11]

    and Hoshen, Y

    Reiss, T., Cohen, N., Horwitz, E., Abutbul, R. and Hoshen, Y. (2022) Anomaly detection requires better representations. In European Conference on Computer Vision, 56--68. Springer

    2022

  4. [12]

    H., Ribeiro, M

    Ribeiro, A. H., Ribeiro, M. H., Paix \ a o, G. M., Oliveira, D. M., Gomes, P. R., Canazart, J. A., Ferreira, M. P., Andersson, C. R., Macfarlane, P. W., Meira Jr, W. et al. (2020) Automatic diagnosis of the 12-lead ecg using a deep neural network. Nature Communications, 11, 1760

    2020

  5. [13]

    and Furht, B

    Salem, O., Guerassimov, A., Mehaoua, A., Marcus, A. and Furht, B. (2014) Anomaly detection in medical wireless sensor networks using svm and linear regression models. Int. J. E Health Medical Commun., 5, 20--45. ://api.semanticscholar.org/CorpusID:1862492

    2014

  6. [14]

    and Komuro, I

    Sawano, S., Kodera, S., Takeuchi, H., Sukeda, I., Katsushika, S. and Komuro, I. (2022) Masked autoencoder-based self-supervised learning for electrocardiograms to detect left ventricular systolic dysfunction. In NeurIPS 2022 Workshop on Learning from Time Series for Health

    2022

  7. [15]

    J., Richter, F., Zhao, S., Vaid, A., Chaudhry, F., De Freitas, J

    Somani, S., Russak, A. J., Richter, F., Zhao, S., Vaid, A., Chaudhry, F., De Freitas, J. K., Naik, N., Miotto, R., Nadkarni, G. N. et al. (2021) Deep learning and the electrocardiogram: review of the current state-of-the-art. EP Europace, 23, 1179--1191

    2021

  8. [16]

    and Pei, D

    Su, Y., Zhao, Y., Niu, C., Liu, R., Sun, W. and Pei, D. (2019) Robust anomaly detection for multivariate time series through stochastic recurrent neural network. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD '19, 2828–...

    2019

  9. [17]

    and Jennings, N

    Tuli, S., Casale, G. and Jennings, N. R. (2022) Tranad: deep transformer networks for anomaly detection in multivariate time series data. Proceedings of the VLDB Endowment, 15, 1201--1214

    2022

  10. [18]

    N., Kaiser,

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, . and Polosukhin, I. (2017) Attention is all you need. Advances in Neural Information Processing Systems, 30

    2017

  11. [19]

    I., Samek, W

    Wagner, P., Strodthoff, N., Bousseljot, R.-D., Kreiseler, D., Lunze, F. I., Samek, W. and Schaeffter, T. (2020) Ptb-xl, a large publicly available electrocardiography dataset. Scientific Data, 7, 1--15

    2020

  12. [20]

    N., Nag, A

    Wang, G., Wang, Q., Iyer, G. N., Nag, A. and John, D. (2023 a ) Unsupervised pre-training using masked autoencoders for ecg analysis. In 2023 IEEE Biomedical Circuits and Systems Conference (BioCAS), 1--5. IEEE

    2023

  13. [21]

    W., Ye, C

    Wang, H., Luo, Z., Yip, J. W., Ye, C. and Zhang, M. (2023 b ) Ecggan: A framework for effective and interpretable electrocardiogram anomaly detection. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 5071--5081

    2023

  14. [22]

    and Long, M

    Xu, J., Wu, H., Wang, J. and Long, M. (2022) Anomaly transformer: Time series anomaly detection with association discrepancy. In International Conference on Learning Representations. ://openreview.net/forum?id=LzQQ89U1qm_

    2022

  15. [23]

    H., Tan, I

    Yaacob, A. H., Tan, I. K., Chien, S. F. and Tan, H. K. (2010) Arima based network anomaly detection. In 2010 Second International Conference on Communication Software and Networks, 205--209

    2010

  16. [24]

    and Zeng, Z

    Yang, S., Lian, C. and Zeng, Z. (2022) Masked autoencoder for ecg representation learning. In 2022 12th International Conference on Information Science and Technology (ICIST), 95--98. IEEE

    2022

  17. [25]

    I., Pan, S., Aggarwal, C

    Zamanzadeh Darban, Z., Webb, G. I., Pan, S., Aggarwal, C. and Salehi, M. (2024) Deep learning for time series anomaly detection: A survey. ACM Computing Surveys, 57, 1--42

    2024

  18. [26]

    Zhang, C., Zhang, C., Song, J., Yi, J. S. K. and Kweon, I. S. (2023) A survey on masked autoencoder for visual self-supervised learning. In IJCAI, 6805--6813

    2023

  19. [27]

    and Huang, Q

    Zhang, H., Liu, W., Shi, J., Chang, S., Wang, H., He, J. and Huang, Q. (2022) Maefe: Masked autoencoders family of electrocardiogram for self-supervised pretraining and transfer learning. IEEE Transactions on Instrumentation and Measurement, 72, 1--15

    2022

  20. [28]

    and Wang, R

    Zheng, Y., Liu, Z., Mo, R., Chen, Z., Zheng, W.-s. and Wang, R. (2022) Task-oriented self-supervised learning for anomaly detection in electroencephalography. In International Conference on Medical Image Computing and Computer-Assisted Intervention, 193--203. Springer

    2022

  21. [29]

    and Zhao, W

    Zhou, Y., Diao, X., Huo, Y., Liu, Y., Fan, X. and Zhao, W. (2023) Masked transformer for electrocardiogram classification. arXiv preprint arXiv:2309.07136

    2023 arXiv

  22. [30]

    R., Cheng, W., Lumezanu, C., ki Cho, D

    Zong, B., Song, Q., Min, M. R., Cheng, W., Lumezanu, C., ki Cho, D. and Chen, H. (2018) Deep autoencoding gaussian mixture model for unsupervised anomaly detection. In International Conference on Learning Representations. ://api.semanticscholar.org/CorpusID:51805340

    2018

  23. [31]

    , " * write output.state after.block = add.period write newline

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.senten...

  24. [32]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.