Automatic Detection of ECG Abnormalities by using an Ensemble of Deep Residual Networks with Attention

Henggui Zhang; Kuanquan Wang; Na Zhao; Qiang Sun; Qince Li; Runnan He; Yang Liu

REVIEW 4 major objections 5 minor 9 references

Automatic Detection of ECG Abnormalities by using an Ensemble of Deep Residual Networks with Attention

T0 review · 4 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read Residual networks with attention reach 0.875 F1 on ECG abnormalities.

desk verdict Reported F1 of 0.875 is plausible but rests on an unspecified test-time windowing rule, so the headline number is not yet reproducible from the paper. read the letter →

arxiv 1908.10088 v1 pith:N3MEFGIA submitted 2019-08-27 cs.LG eess.SPstat.ML

Yang Liu , Runnan He , Kuanquan Wang , Qince Li , Qiang Sun , Na Zhao , Henggui Zhang This is my paper · ORCID

classification cs.LGeess.SPstat.ML

keywords electrocardiogram12-leadECGdeepresidualnetworkattentionmechanismensemblemodelmulti-labelclassificationdataaugmentationF1score

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 reports a computer pipeline for spotting nine types of abnormality in 12-lead ECG recordings. The pipeline denoises the signal, randomly pads or truncates every recording to 30 seconds, augments and re-balances the classes, and then trains an ensemble of deep residual convolutional networks whose final layer is an attention-weighted summary of local waveform features. On the hidden test set of the First China ECG Intelligent Competition, the ensemble achieves an overall F1 of 0.875, computed as the average of the nine per-class F1 scores. If that score holds, the approach is a plausible triage tool that could flag suspected cases for human review rather than replace a cardiologist.

What carries the argument

The carrying mechanism is a residual convolutional network with an attention layer, applied to 30-second slices of 12-lead ECG. The local-feature stage passes the raw signal through a 1D convolution followed by seven residual modules, each consisting of batch normalization, dropout, ReLU, and 1D convolution with an additive skip connection, with max-pooling after each module. The attention stage turns the resulting sequence of local feature vectors into a single global vector by learning a weight for each local vector and summing them; a fully-connected layer with per-class sigmoid outputs then makes independent predictions for each of the nine labels. Attention is what lets the model concentrate on the short episodes in a recording that actually carry a diagnosis, while the random padding/truncating scheme is what keeps the network from latching onto recording length or padding position.

What would settle it

One decisive check: rescore each hidden-test recording with a different randomly chosen 30-second window and compare per-class F1; if scores for paroxysmal classes such as PVC or PAC drop, the 0.875 result depends on which window the model saw, not on a stable signal of the abnormality.

Watch

Extended reading notes

Core claim

The central claim is that an ensemble of residual networks with attention can classify nine ECG abnormality classes with a macro-averaged F1 of 0.875 on an unseen test set. Per-class results are above 0.9 for atrial fibrillation, first-degree atrioventricular block, complete right bundle branch block, premature ventricular contraction, and premature atrial contraction, and near 0.75 for left anterior fascicular block, early repolarization pattern changes, and T-wave changes. The authors attribute the result to two design choices: random padding/truncating to a 30-second target length, which augments and balances the data while suppressing overfitting to where padding sits, and a model that first learns local features through seven residual convolutional modules and then weights those features with an attention distribution before classification. The ensemble averages probabilistic predictions from four training pipelines, one of which mixes in an external dataset and another of which retrains balanced models on imbalanced data.

Load-bearing premise

The load-bearing premise is that a randomly selected 30-second slice of a longer ECG recording contains every abnormality that a clinician would need to classify; the paper reports no measurement of how often truncation discards the relevant episode.

Editorial extensions

If this is right

If the 0.875 macro-F1 transfers to other patient cohorts, the pipeline can pre-screen 12-lead ECGs and route high-confidence abnormal recordings to a cardiologist before full manual review.
For the five classes with per-class F1 above 0.9, automated flags could be usable as decision support without per-case recalibration.
The three low-F1 classes (LAFB, ER, TWC) identify where the method needs more training data or additional features, since the paper attributes their weakness to relatively few samples.
Because the architecture is an ensemble of independently trained pipelines, adding pipelines with different preprocessing is a direct route to incremental accuracy gains.
The random padding/truncation augmentation means the model is intentionally insensitive to the temporal position of a 30-second window, which is a testable assumption about whether short windows are enough for ECG diagnosis.

Reading between the lines

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

Editorial inference: if the random 30-second window sometimes cuts off a transient abnormality, the reported F1 could overstate what the model does on a full recording; re-scoring with several windows per test recording would quantify this.
Editorial inference: the attention weights learned by the network are a ready-made saliency map over the ECG trace, but the paper does not check whether those weights align with where clinicians locate abnormalities.
Editorial inference: the two-stage balance-then-imbalance training used in one pipeline suggests a general strategy for imbalanced medical datasets, but its contribution is not isolated by an ablation.
Editorial inference: because each class uses an independent sigmoid, lowering the decision threshold would trade precision for recall and could recover F1 on LAFB, ER, and TWC.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The paper reports a strong result: F1 of 0.875 on a nine-class, multi-label 12-lead ECG classification task from an external hidden test set. That number is the whole ballgame, and it is both plausible and under-specified. The random padding/truncating augmentation is a simple, sensible idea to both regularize and balance training data, and the two-stage balance-then-retrain-with-imbalanced-data in pipeline 3 is a neat trick. The architecture itself is not new—residual blocks plus attention—but the specific combination with the random windowing strategy is a legitimate new application. The evaluation is against a held-out test set, so the headline score is not a re-statement of the training objective, and the F1 definitions are clear and internally consistent.

The soft spots are real, and one is load-bearing. Nowhere in Sections 3.3 or 4 does the paper say how test recordings longer than 30 seconds are reduced to 30 seconds at inference. For paroxysmal classes like PVC, PAC, and ER, a single 30-second window may not contain the abnormal beat; if the crop is random, a single test recording could receive different predictions on different runs. The training-time random window is described, but the inference-time protocol is not. That means the reported 0.875 depends on an unstated, possibly stochastic rule, and the reader cannot verify the central quantitative claim from the text alone. This is a reporting gap, not evidence the result is wrong, but it is a serious one.

Beyond that, the paper has no confidence intervals, no comparison to any baseline method (not even a single-pipeline version), and no ablation supporting the Conclusion's claim that the ensemble beat each pipeline individually. No code or data are released. The citation pattern is fine—references are standard, the one self-citation is relevant. The paper takes a small step beyond prior work, but the incremental novelty is acknowledged by the authors' own references showing prior CNN, residual, and RNN ECG classifiers.

Overall, this looks like a competent competition write-up with an incomplete technical description. It deserves a serious referee—the question of whether the reported result holds up is important for applied ECG deep learning—but it needs major revision: spell out the test-time windowing, add baselines and ablations, and ideally release code or predictions so the F1 can be checked. I would not cite it as a trustworthy baseline until that is done.

Send it to peer review, but expect heavy revision.

Referee Report

4 major / 5 minor

Summary. The paper proposes an automatic ECG abnormality detection pipeline for 12-lead recordings, combining preprocessing (baseline wander removal, wavelet denoising, length unification, data augmentation and class balancing), a deep residual convolutional network with attention, and an ensemble of models trained under four different pipelines. The method is evaluated on the hidden test set of the First China ECG Intelligent Competition using class-wise F1 and their macro-average, with a reported overall F1 of 0.875 across nine classes. The central claim is that this ensemble of attention-augmented residual networks yields strong multi-label classification performance on a realistic ECG dataset.

Significance. If the reported F1 of 0.875 is reproducible and robust, the result would be of practical value for automated ECG screening, especially because the problem is multi-label, the recordings vary in length up to 90 s, and some abnormalities are paroxysmal. The manuscript is commendable for providing explicit formulas for the F1 metric (Eqs. 2-9), for evaluating on an external hidden test set rather than a held-out portion of the training data, and for describing the training pipelines in enough detail to be partially replicable. However, the significance is substantially limited by the absence of baseline comparisons, the lack of any uncertainty quantification, and the underspecification of the test-time windowing rule, all of which prevent the reader from calibrating or verifying the headline number. The paper also claims that the ensemble is superior to individual pipelines, but it does not report the supporting per-pipeline results.

major comments (4)

[§3.1, §3.3, §4] The inference-time windowing rule for test recordings longer than 30 s is never specified. Section 3.1 describes random padding/truncating as a training-time augmentation, and Section 3.3 explains how this creates training samples, but Section 4 reports F1 = 0.875 on the entire test set without stating whether a fixed 30-s window is cut from the beginning, a random window is sampled once, or multiple windows are aggregated. Because the recording-level labels of paroxysmal classes (PVC, PAC, ER) may appear only in a short segment, the reported macro-F1 depends on this underspecified choice. Please state the exact test-time protocol and demonstrate that the result is stable under different window positions, for example by reporting variation across several random windows or a deterministic aggregation rule.
[Table 1 and §4] No uncertainty or significance estimates are reported for the macro-F1. With 8,500 test recordings and nine classes of very different prevalence (e.g., LAFB, ER, TWC at F1 ≈ 0.73-0.76), the single number 0.875 gives no indication of whether differences between models or between classes are meaningful. Add bootstrap confidence intervals for each class-level F1 and for the macro-F1, or state the official competition evaluation protocol if it defines a fixed scoring rule.
[§4] No comparison with any baseline method is provided. The manuscript reports only the proposed method's F1, so the reader cannot determine whether the 0.875 score represents a genuine advance over, for example, a standard CNN, an LSTM-based model, or a simpler non-attention residual network. A comparison with at least one previously published ECG classifier on the same competition test set is needed to support the claimed promising performance.
[§5] The conclusion states that the proposed ensemble 'achieved better results than each of the separated pipelines,' but no per-pipeline F1 scores or ensemble ablations are reported anywhere in the paper. This claim is therefore not supported by the presented evidence. Please provide the per-pipeline results or remove the superiority claim.

minor comments (5)

[§3.1, Eq. (1)] Equation (1) does not state the units of N, and the text gives conflicting cut-off frequencies: the stated lower bound is 0.5 Hz, but N = 500 gives f_co = 0.443 Hz, which is below that bound. Please clarify the intended window sizes and the corresponding cut-off frequencies.
[§3.2] The text says the feature map length becomes 1/27 of the input length after the local feature learning stage, but the architecture has seven max-pooling layers; unless the pooling factors are not 2, this should be 1/128. Please justify or correct this factor.
[§3.3 and Fig. 2] The description of pipeline 4 is incomplete: it says the input is normalized to zero mean and unit variance but does not specify whether normalization is applied per-lead, per-recording, or across the dataset, nor how this interacts with the denoising used in the other pipelines.
[Table 1] The table reports F1 scores but not the number of test samples or the prevalence of each class. Adding per-class support counts would help interpret the low F1 values for LAFB, ER, and TWC.
[General] The paper would benefit from a summary of the official competition scoring rule and a clear statement of whether the reported 0.875 is the exact official score or a recomputed value, since the metric definition in Eqs. (2)-(9) may differ from the competition's own protocol.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the reported F1=0.875 is an evaluation on an external hidden test set, and no prediction reduces to a fitted input or self-citation.

full rationale

The paper's central result is an ensemble classifier scored on the hidden test set of the First China ECG Intelligent Competition. The F1 formulas in Equations (2)-(9) define the metric from predicted label sets f(x_i) and annotated label sets Y_i, and this scoring is external to the training procedure. No parameter is fitted to test labels: the class-balancing and two-stage retraining in Section 3.3 uses the class and length distributions of the training data, plus the stated assumption that the test set shares the same distribution, but it never uses test labels. The random padding/truncating window selection is a training-time data augmentation strategy; even if the test-time windowing rule is not fully specified, that is a reproducibility or correctness concern rather than a circular derivation. The only self-citation, reference [7], is cited merely as an example of convolutional-network ECG work and is not load-bearing for the proposed residual-attention ensemble, which is compared against an independent benchmark. Therefore, no step in the paper's derivation chain reduces by construction to its own inputs.

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

The paper introduces no new physical entities. Its central result rests on several unvalidated domain assumptions and numerous hand-chosen hyperparameters; the main free parameters are the input length and network architecture choices, plus unreported training details.

free parameters (5)

Target signal length = 15000 samples (30s)
Chosen because more than 90% of recordings are no longer than 30s; determines how much signal is truncated or padded.
Baseline removal window size = 250 (pipeline 1), 500 (pipelines 2-4)
Sets the cutoff frequency for baseline wander removal; chosen by hand without reported sensitivity analysis.
Kernel size and network width schedule = kernel size 16; kernel number starts at 16 and grows by 16 every two residual modules
Architecture hyperparameters selected without reported search or ablation.
Number of residual modules = 7
Architecture depth chosen by hand.
Training hyperparameters = not reported
Optimizer, learning rate, batch size, epochs, and dropout rates are not given in the paper, yet they affect the final F1.

assumptions (6)

domain assumption Residual convolutional modules improve feature learning and avoid degradation in deep networks.
Relies on He et al. (2016) empirical finding, cited as reference [6].
domain assumption An attention mechanism can summarize the clinically relevant parts of a long ECG recording.
Assumed in Section 3.2; no validation of the attention weights is provided.
domain assumption Wavelet denoising with 5-level db4 and soft-thresholding removes powerline and muscle noise without removing diagnostic signal.
Used in Section 3.1; no quantitative verification on this dataset.
domain assumption Truncating recordings longer than 30 seconds to a random 30-second window preserves the abnormalities used for labeling.
Adopted in Section 3.1 to unify lengths; no analysis of information loss.
domain assumption The training and test sets come from the same distribution, and the test labels are correct.
Assumed in Section 3.3 when using the original imbalanced distribution for retraining and when reporting final F1.
domain assumption CPSC 2018 recordings can be mixed with the competition training data without harming performance.
Used in pipeline 2; no domain-shift analysis.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Automatic Detection of ECG Abnormalities by using an Ensemble of Deep Residual Networks with Attention." pith.science (2026). https://pith.science/paper/N3MEFGIA

@misc{pith2026190810088,
  author       = {Pith},
  title        = {Pith review of: Automatic Detection of ECG Abnormalities by using an Ensemble of Deep Residual Networks with Attention},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N3MEFGIA}},
  note         = {Machine review of arXiv:1908.10088}
}

read the original abstract

Heart disease is one of the most common diseases causing morbidity and mortality. Electrocardiogram (ECG) has been widely used for diagnosing heart diseases for its simplicity and non-invasive property. Automatic ECG analyzing technologies are expected to reduce human working load and increase diagnostic efficacy. However, there are still some challenges to be addressed for achieving this goal. In this study, we develop an algorithm to identify multiple abnormalities from 12-lead ECG recordings. In the algorithm pipeline, several preprocessing methods are firstly applied on the ECG data for denoising, augmentation and balancing recording numbers of variant classes. In consideration of efficiency and consistency of data length, the recordings are padded or truncated into a medium length, where the padding/truncating time windows are selected randomly to sup-press overfitting. Then, the ECGs are used to train deep neural network (DNN) models with a novel structure that combines a deep residual network with an attention mechanism. Finally, an ensemble model is built based on these trained models to make predictions on the test data set. Our method is evaluated based on the test set of the First China ECG Intelligent Competition dataset by using the F1 metric that is regarded as the harmonic mean between the precision and recall. The resultant overall F1 score of the algorithm is 0.875, showing a promising performance and potential for practical use.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

9 extracted references · 9 canonical work pages

[1]

K., Chacko, A.: ECG signal enhancement using S-Transform,” IEEE Trans

Ari, S., Das, M. K., Chacko, A.: ECG signal enhancement using S-Transform,” IEEE Trans. Biomed. Eng. 43 (6) , 649-660 (2013)

work page 2013
[2]

De Chazal, P., Reilly, R.B.: A patient-adapting heartbeat classifier using ECG morphology and heartbeat interval features, IEEE Trans. Biomed. Eng. 53 (12), 2535–2543 (2006)

work page 2006
[3]

IEEE Trans

Yang, H., Kan, C., Liu, G., Chen, Y.: Spatiotemporal differentiation of myocardial infarc- tions. IEEE Trans. Autom. Sci. Eng. 10 (4), 938–947 (2013)

work page 2013
[4]

IEEE Trans

Dima, S.-M., Panagiotou, C., Mazomenos, E.B., Rosengarten, J.A., Maharatna, K., Gialelis, J.V., Curzen, N., Morgan, J.: On the detection of myocadial scar based on ECG/VCG anal- ysis. IEEE Trans. Biomed. Eng. 60 (12), 3399–3409 (2013)

work page 2013
[5]

Chou, K.T.: I ntegration of independent component analysis and neural networks for ECG beat classification

Yu, S.N. Chou, K.T.: I ntegration of independent component analysis and neural networks for ECG beat classification. Expert Syst. Appl 34, 2841–2846 (2008)

work page 2008
[6]

In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770 –778. IEEE, Las Vegas, NV, USA (2016)

work page 2016
[7]

Frontiers in physiology 9, 1206 (2018)

He, R., Wang, K., Zhao, N., Liu, Y., Yuan, Y., Li, Q., Zhang, H.: Automatic detection of atrial fibrillation based on continuous wavelet transform and 2D convolutional neural net- works. Frontiers in physiology 9, 1206 (2018)

work page 2018
[8]

L., Ng, E

Oh, S. L., Ng, E. Y., San Tan, R., Acharya, U. R.: Automated diagnosis of arrhythmia using combination of CNN and LSTM techniques with variable length heart beats. Computers in biology and medicine 102, 278-287 (2018)

work page 2018

Show all 9 references

[9]

Donoho, D.L.: De-noising by soft-thresholding, IEEE Trans. Inform. The. 41 (3), 613–627 (1995)

1995

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

K., Chacko, A.: ECG signal enhancement using S-Transform,” IEEE Trans

Ari, S., Das, M. K., Chacko, A.: ECG signal enhancement using S-Transform,” IEEE Trans. Biomed. Eng. 43 (6) , 649-660 (2013)

work page 2013

[2] [2]

De Chazal, P., Reilly, R.B.: A patient-adapting heartbeat classifier using ECG morphology and heartbeat interval features, IEEE Trans. Biomed. Eng. 53 (12), 2535–2543 (2006)

work page 2006

[3] [3]

IEEE Trans

Yang, H., Kan, C., Liu, G., Chen, Y.: Spatiotemporal differentiation of myocardial infarc- tions. IEEE Trans. Autom. Sci. Eng. 10 (4), 938–947 (2013)

work page 2013

[4] [4]

IEEE Trans

Dima, S.-M., Panagiotou, C., Mazomenos, E.B., Rosengarten, J.A., Maharatna, K., Gialelis, J.V., Curzen, N., Morgan, J.: On the detection of myocadial scar based on ECG/VCG anal- ysis. IEEE Trans. Biomed. Eng. 60 (12), 3399–3409 (2013)

work page 2013

[5] [5]

Chou, K.T.: I ntegration of independent component analysis and neural networks for ECG beat classification

Yu, S.N. Chou, K.T.: I ntegration of independent component analysis and neural networks for ECG beat classification. Expert Syst. Appl 34, 2841–2846 (2008)

work page 2008

[6] [6]

In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770 –778. IEEE, Las Vegas, NV, USA (2016)

work page 2016

[7] [7]

Frontiers in physiology 9, 1206 (2018)

He, R., Wang, K., Zhao, N., Liu, Y., Yuan, Y., Li, Q., Zhang, H.: Automatic detection of atrial fibrillation based on continuous wavelet transform and 2D convolutional neural net- works. Frontiers in physiology 9, 1206 (2018)

work page 2018

[8] [8]

L., Ng, E

Oh, S. L., Ng, E. Y., San Tan, R., Acharya, U. R.: Automated diagnosis of arrhythmia using combination of CNN and LSTM techniques with variable length heart beats. Computers in biology and medicine 102, 278-287 (2018)

work page 2018