Pith. sign in

REVIEW 2 major objections 4 minor 64 references

A guideline-grounded global evaluation shows that most gradient-based explanation methods transferred from computer vision systematically follow ECG signal amplitude rather than clinically diagnostic regions, and that sign-adjusted variants

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-07-31 23:13 UTC pith:QNPYXIDZ

load-bearing objection A transparent, reusable benchmark that convincingly demonstrates amplitude bias in gradient-based ECG attributions; the abstract just overstates the faithfulness claim. the 2 major comments →

arxiv 2607.24035 v1 pith:QNPYXIDZ submitted 2026-07-27 cs.CY cs.LG

Beyond Local Inspection: Global, Guideline-Grounded Evaluation of Post-hoc XAI Methods for ECG Classification

Nils Gumpfer , Michael Guckert , Samuel Sossalla , Birgit A{\ss}mus , Jennifer Hannig This is my paper
classification cs.CY cs.LG
keywords explainable AIECG classificationpost-hoc attributionclinical plausibilityamplitude biasguideline-based evaluationnormalized coveragetime-series XAI
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that popular gradient-based post-hoc explanation methods, when applied to ECG classifiers without domain adaptation, are systematically driven by signal amplitude rather than by clinically diagnostic patterns, and that this failure is invisible in single-sample heatmaps. To show this, it aggregates explanations across thousands of heartbeats and compares their coverage of cardiologist-defined diagnostic regions (PR interval, ST segment, QRS morphology) on four conditions spanning low- and high-amplitude criteria. The evidence: input-multiplied methods such as Integrated Gradients and LRP-ε reach Spearman correlations with absolute signal amplitude up to 0.69 and assign as little as 4.6% of relevance to the ischemia-defining ST segment; nine of the thirteen methods score below chance on at least one condition. Sign-adjusted variants, which multiply gradients by only the sign of the input, remain above chance on all four pathologies, with LRP-ε/SIGN reaching the highest worst-case normalized coverage (41.5%). A sympathetic reader cares because choosing an explanation method for clinical AI is currently guided by convention and visual plausibility, and the paper offers a quantitative, guideline-grounded way to detect systematic attribution failures before deployment.

Core claim

The paper claims that ECG explanation quality should be judged globally, against clinical guidelines, and that on this test most gradient-based attribution methods fail. Comparing 13 methods on the same four trained CNNs and the same true-positive predictions, it shows that input-multiplied and baseline-subtracted methods (Gradient×Input, Integrated Gradients, DeepSHAP, GradSHAP, LRP-ε, SmoothGrad×Input) concentrate relevance on the high-amplitude R-peak and T-wave regardless of pathology, while sign-adjusted versions keep relevance on the low-amplitude PR interval and ST segment that define AV block and ischemia. Nine of thirteen methods fall below a random baseline on at least one conditio

What carries the argument

The framework's central mechanism is a five-step global aggregation pipeline that turns local attribution maps into population-level relevance histograms over R-peak-aligned 500-timestep heartbeats, then scores them against guideline-derived binary reference masks (PR interval for AV block, ST segment for ischemia, QRS complex for both bundle branch blocks). Two metrics carry the argument: normalized coverage (NCov), a skill score expressing how much of the positive relevance falls in the masked region relative to a random baseline, and its minimum across pathologies (NCovmin), which ranks methods by their weakest condition rather than their average. A Spearman correlation between absolute r

Load-bearing premise

The central load-bearing premise is that the cardiologist-derived reference masks are the correct target and that the predictively valid CNNs, applied only to true-positive cases, act as a pattern proxy — meaning coverage scores plausibility, not faithfulness, and the paper explicitly states it did not test the method-model linkage directly.

What would settle it

Train an ischemia classifier that is deliberately prevented from using the ST segment (e.g., by masking that region during training) yet still reaches high accuracy through other features, and run LRP-ε/SIGN through the pipeline. If it still shows high normalized coverage of the ST mask, then the coverage score reflects the method's bias toward the mask rather than clinically meaningful model support, and the ranking among methods would not be evidence of clinical plausibility.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Global, guideline-grounded evaluation can expose systematic attribution failures that local heatmap inspection cannot, so studies relying on visual plausibility alone are insufficient evidence for clinical trust.
  • Amplitude-multiplied and baseline-subtracted methods (Integrated Gradients, DeepSHAP, LRP-ε, SmoothGrad×Input) should not be assumed reliable for low-amplitude ECG criteria; their high average coverage is driven by bundle branch block cases where diagnostic regions coincide with high-amplitude waveforms.
  • Methods should be selected by worst-case normalized coverage across conditions rather than mean coverage, because a method is deployed once and then meets whichever pathology appears.
  • The SIGN adjustment restores above-chance attribution on all four pathologies for all three SIGN variants, with LRP-ε/SIGN the best; this positions sign-adjusted gradients as a low-cost default for time-series XAI.
  • The framework is method- and model-agnostic, so the same mask-and-coverage pipeline can audit different models and architectures, or be extended to other physiological time series with guideline-defined regions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The amplitude-bias mechanism likely generalizes beyond ECG: any physiological time series whose diagnostic features sit in low-amplitude segments (e.g., EEG event-related potentials or PPG dicrotic notch timing) would be vulnerable to the same input-multiplication bias, and sign-adjusted gradients are a natural first test.
  • Because the paper uses the trained model as a fixed pattern proxy and does not test faithfulness, a high-coverage method could still be exploiting a shortcut shared with the mask; combining this framework with model-randomization or feature-deletion tests would separate method bias from model behavior.
  • A clinician-facing implication left implicit: presenting global guideline-coverage scores alongside local heatmaps would give cardiologists a decision-relevant summary of whether an explanation method tracks the region they actually read; structured clinician studies could test whether coverage scores change trust or diagnostic agreement.
  • The mask-based metric is blind to morphological shape within the QRS complex, as the paper notes; extending it to concept-level attribution (e.g., concept relevance propagation) would let coverage register whether relevance follows the notched or slurred morphology that distinguishes the bundle branch blocks.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper proposes a global, guideline-grounded evaluation framework for post-hoc XAI methods in ECG classification. Using four binary CNN classifiers trained on PTB-XL, the authors apply 13 gradient-based attribution methods to true-positive test predictions, extract and align beats, and aggregate relevance across beats and leads. They compare these aggregated attributions against guideline-derived reference masks (PR interval for AV block, ST segment for ischemia, QRS complex for RBBB/LBBB) using coverage (Eq. 4), normalized coverage relative to a random baseline (Eq. 5), and worst-case normalized coverage over pathologies (Eq. 6). They also report Spearman correlations between absolute relevance and absolute input amplitude. The main results are that input-multiplied or baseline-dependent methods (Integrated Gradients, LRP-ε, DeepSHAP, SmoothGrad×Input) strongly correlate with signal amplitude (mean SCC up to 0.69) and have negative normalized coverage on low-amplitude diagnostic regions (e.g., LRP-ε: 4.6% coverage on the ischemia ST segment), while SIGN-adjusted variants maintain positive worst-case coverage across all four pathologies. The paper concludes that methods transferred from computer vision systematically overlook diagnostically decisive low-amplitude regions and recommends SIGN-based adaptations.

Significance. The framework is a genuinely useful contribution to time-series XAI evaluation: it is open-source, reusable, uses expert-defined reference masks, aggregates over many heartbeats to reveal systematic patterns, and reports worst-case reliability rather than only averages. The bootstrap confidence intervals and mask-boundary sensitivity analysis strengthen the internal validity of the comparisons. If the central assumption holds—that the trained CNNs rely on the guideline-defined regions when making true-positive predictions—the finding that many popular attribution methods are amplitude-dominated in ECG would be clinically important and would motivate domain-adapted XAI. The authors are transparent that they measure clinical plausibility (guideline agreement), not faithfulness, and they acknowledge this in the Methods and Limitations. The significance is therefore conditional on the pattern-proxy assumption, which is plausible but not directly tested in this manuscript.

major comments (2)
  1. [Abstract and Conclusion] The abstract and conclusion use causal language that goes beyond the measured construct. The abstract states that methods "overlook diagnostically decisive low-amplitude regions" and concludes a "systematic failure" of CV-transferred methods. However, the Methods section ("Clinical Plausibility vs. Faithfulness") explicitly states: "we did not test that linkage directly" and that coverage measures clinical plausibility, not faithfulness. A method can have low ST-segment coverage either because it fails to expose the model's reasoning or because the model itself does not rely on the ST segment. The same-model argument (LRP-SIGN 63.8% vs LRP-ε 4.6% on ISCH) only proves that at least one of the two is not perfectly faithful; it does not identify which. The abstract and conclusion should be rephrased to say that these methods show "low agreement with guideline-defined diagnostic regions" or
  2. [Methods — Clinical Plausibility vs. Faithfulness; Discussion] The pattern-proxy assumption is load-bearing but untested. The paper relies on three indirect justifications: (i) the models are predictively valid (AUROC 0.93–0.99), (ii) the analysis is confined to true positives, and (iii) prior lead-level evidence on the same architecture/pathologies. These do not establish that the CNNs use the guideline-defined temporal regions. The statement that "one model cannot account for both" (LRP-ε/SIGN vs LRP-ε) is insufficient: two unfaithful methods can disagree arbitrarily on a fixed model. If the CNNs actually rely on R-peak/T-wave amplitude or other shortcuts, then LRP-ε could be more faithful than SIGN despite lower coverage. The paper should either add a targeted test (e.g., perturbing/ablating the guideline region and measuring the model's output change) or explicitly frame the results as "guideline agreement," not as evidence about the model's dec
minor comments (4)
  1. [Eq. (3)] The definition of Spearman's correlation is unnecessarily convoluted. A standard formulation using ranks would be clearer and shorter.
  2. [Methods — Pattern Analysis] The threshold of 0.05 for discarding positive relevance values is a free parameter. The paper does not report sensitivity to this threshold. Since the results depend on it, a small sensitivity analysis (e.g., thresholds 0.01, 0.1) would be reassuring.
  3. [Coverage metric / Limitations] The coverage metric pools relevance across all 12 leads and applies the same temporal mask to every lead. As the authors note, relevance placed in the correct temporal region but in a clinically uninformative lead is still credited. This is acknowledged, but the abstract's phrase "diagnostically decisive regions" could be more precise: the masks are temporal, not lead-specific, and the paper evaluates "temporal agreement pooled over leads."
  4. [Figures 1–2] The figures are dense and difficult to read in the manuscript layout. The 16-panel histograms in Figure 2 would benefit from larger panels or a zoomed-in view for at least one representative lead.

Circularity Check

0 steps flagged

No derivation reduces to its inputs; disclosed self-citations and an explicitly untested pattern-proxy assumption are the only caveats, not circular reductions.

full rationale

The paper's central results are empirical: four CNNs trained on PTB-XL, 13 attribution methods, and coverage/normalized-coverage/worst-case metrics (Eqs. 4-6) computed from relevance maps and guideline masks. No parameter is fitted to the target outcome and no method's score is an identity with its definition; the high SCC of input-multiplied methods is acknowledged as partly true by construction and the paper rests its conclusion on the coverage divergence instead. The SIGN recommendation is based on the same open pipeline as all methods and the self-origin is disclosed in Competing Interests. The main interpretive step is the pattern-proxy assumption, for which the paper states 'we did not test that linkage directly' and cites the authors' own prior work for lead-level agreement; this is an explicit validity limitation and a mild self-citation dependency, but it does not make the numerical comparison reduce to the guideline masks or to the authors' prior conclusions. The framework is self-contained against external data and clinical guidelines, so no circular step is identified. Score 2 reflects the minor self-citation/pattern-proxy caveat rather than any derivation-level circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central evaluation depends on expert-defined temporal masks and the assumption that a predictively valid model's true-positive explanations should align with guideline regions. These are domain assumptions the authors disclose; the main hand-chosen numeric inputs are the mask boundaries and the 0.05 relevance threshold, tested only partially (mask shifts) or not at all (threshold).

free parameters (3)
  • Reference mask temporal boundaries = Clinical segments in standardized 500-step beat (exact boundaries not given in paper)
    Central to the coverage metric; hand-set temporal segments defined by two cardiologists and verified against median beats; sensitivity analysis around ±20 steps does not eliminate dependence on the central choice.
  • Positive-relevance filtering threshold = 0.05
    Normalized positive relevance below 0.05 is discarded; affects all coverage and histogram results; no sensitivity analysis for this threshold is reported.
  • Number of extracted beats per recording = 2
    Fixed sampling-density choice for the aggregation pipeline; reasonable, but a free protocol choice.
axioms (5)
  • domain assumption Guideline-defined diagnostic regions are an appropriate ground truth for explanation quality.
    Reference Patterns: the published guidelines (29-31) are treated as the gold standard for what a trustworthy explanation should emphasize.
  • domain assumption Predictively valid CNNs on true-positive predictions are a pattern proxy; method rankings on a fixed model are meaningful.
    Methods, Clinical Plausibility vs. Faithfulness: all methods share the same fixed model per pathology, and coverage differences are attributed to the methods.
  • domain assumption Only positive relevance supports the predicted class; negative relevance carries no guideline support.
    Pattern Analysis: 'Restricting the analysis to positive relevance allowed us to isolate features supporting the predicted class'; this may under-credit methods that encode evidence in negative relevance.
  • domain assumption Temporal-only masks applied across all leads sufficiently represent diagnostic criteria.
    Limitations: masks encode only the temporal component of the criteria and apply across all leads; lead focus is only visible in histograms, not in the coverage metric.
  • standard math Spearman rank correlation is the appropriate measure of amplitude dependence.
    Equation (3); a standard non-parametric measure of monotone dependence, appropriate for the stated goal but not a causal measure.

pith-pipeline@v1.3.0-alltime-deepseek · 21141 in / 17033 out tokens · 151568 ms · 2026-07-31T23:13:01.205669+00:00 · methodology

0 comments
read the original abstract

Explainable AI (XAI) is used to assess whether artificial intelligence models rely on meaningful patterns, yet explanations that appear plausible for individual predictions may systematically misrepresent model behavior. This is particularly problematic in medicine, where models may rely on irrelevant signal characteristics rather than disease-specific patterns without being recognizable. We address this challenge using electrocardiogram (ECG) data, for which clinical guidelines provide explicit knowledge about diagnostically relevant signal regions. We introduce a global, guideline-grounded framework that aggregates explanations across heartbeats to evaluate them against clinically defined regions of interest. Using four binary classifiers trained on PTB-XL, we assess 13 gradient-based methods across two categories of patterns: low-amplitude segments and high-amplitude QRS morphology. Our results reveal a systematic failure of methods transferred from computer vision. Their explanations often follow signal amplitude rather than clinical relevance, with mean Spearman correlations up to 0.69, leading them to overlook diagnostically decisive low-amplitude regions. For ischemia, LRP-$\epsilon$ assigns only 4.6% of relevance to the ST segment, compared with 63.8% for LRP-SIGN. Nine of 13 methods fall below chance for at least one condition, indicating inconsistent reliability across patterns. These findings show that global, domain-grounded evaluation can uncover systematic explanation failures not obvious from sample-level heatmaps.

Figures

Figures reproduced from arXiv: 2607.24035 by Birgit A{\ss}mus, Jennifer Hannig, Michael Guckert, Nils Gumpfer, Samuel Sossalla.

Figure 1
Figure 1. Figure 1 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Guideline-informed reference masks and the global attribution patterns evaluated against them. Panel a marks the regions that diagnostic guidelines29–31 designate for each pathology, namely PR-interval prolongation for AV-block, ST-segment deviation for ischemia, and QRS morphology in V1-V2 and V5-V6 for the bundle branch blocks. The masks encode only the temporal extent of these criteria and apply across … view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

64 extracted references · 21 canonical work pages

  1. [1]

    & Topol, E

    Rajpurkar, P., Chen, E., Banerjee, O. & Topol, E. J. AI in health and medicine.Nat. Medicine28, 31–38, DOI: 10.1038/s41591-021-01614-0 (2022)

    doi:10.1038/s41591-021-01614-0 2022

  2. [2]

    & Kaynak, O

    Jiang, Y ., Li, X., Luo, H., Yin, S. & Kaynak, O. Quo vadis artificial intelligence?Discov. Artif. Intell.2, DOI: 10.1007/s44163-022-00022-8 (2022)

    doi:10.1007/s44163-022-00022-8 2022

  3. [3]

    E., Attia, Z

    Carter, R. E., Attia, Z. I., López-Jiménez, F. & Friedman, P. A. Pragmatic considerations for fostering reproducible research in artificial intelligence.NPJ Digit. Medicine2, DOI: 10.1038/s41746-019-0120-2 (2019)

    work page doi:10.1038/s41746-019-0120-2 2019

  4. [4]

    Asan, O., Bayrak, A. E. & Choudhury, A. Artificial intelligence and human trust in healthcare: Focus on clinicians.J. Med. Internet Res.22, e15154, DOI: 10.2196/15154 (2020)

    doi:10.2196/15154 2020

  5. [5]

    Is AI leading to a reproducibility crisis in science?Nature624, 22–25, DOI: 10.1038/d41586-023-03817-6 (2023)

    Ball, P. Is AI leading to a reproducibility crisis in science?Nature624, 22–25, DOI: 10.1038/d41586-023-03817-6 (2023)

    work page doi:10.1038/d41586-023-03817-6 2023

  6. [6]

    Commun.10, 1096, DOI: 10.1038/s41467-019-08987-4 (2019)

    Lapuschkin, S.et al.Unmasking clever hans predictors and assessing what machines really learn.Nat. Commun.10, 1096, DOI: 10.1038/s41467-019-08987-4 (2019)

    doi:10.1038/s41467-019-08987-4 2019

  7. [7]

    Y ., Novoa, R

    Narla, A., Kuprel, B., Sarin, K. Y ., Novoa, R. A. & Ko, J. Automated classification of skin lesions: from pixels to practice. J. Investig. Dermatol.138, 2108–2110, DOI: 10.1016/j.jid.2018.06.175 (2018). 8.Hall, J. & Guyton, A.The Normal Electrocardiogram, 121–127 (Saunders/Elsevier, 2011)

    work page doi:10.1016/j.jid.2018.06.175 2018

  8. [9]

    Manimaran, G.et al.Explainable deep learning based techniques for ecg-based heart disease classification: A systematic literature review and future direction.Comput. Biol. Medicine199, 111324, DOI: 10.1016/j.compbiomed.2025.111324 (2025)

    arXiv 2025

  9. [10]

    Attia, Z. I.et al.An artificial intelligence-enabled ECG algorithm for the identification of patients with atrial fibrillation during sinus rhythm: a retrospective analysis of outcome prediction.Lancet394, 861–867, DOI: 10.1016/s0140-6736(19) 31721-0 (2019)

    work page doi:10.1016/s0140-6736(19 2019

  10. [11]

    Y .et al.Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network.Nat

    Hannun, A. Y .et al.Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network.Nat. Medicine25, 65–69, DOI: 10.1038/s41591-018-0268-3 (2019)

    doi:10.1038/s41591-018-0268-3 2019

  11. [12]

    & Strodthoff, C

    Strodthoff, N. & Strodthoff, C. Detecting and interpreting myocardial infarction using fully convolutional neural networks. Physiol. Meas.40, 015001, DOI: 10.1088/1361-6579/aaf34d (2019)

    work page doi:10.1088/1361-6579/aaf34d 2019

  12. [13]

    & Guckert, M

    Gumpfer, N., Grün, D., Hannig, J., Keller, T. & Guckert, M. Detecting myocardial scar using electrocardiogram data and deep neural networks.Biol. Chem.402, 911–923, DOI: 10.1515/hsz-2020-0169 (2020). 14/17

    work page doi:10.1515/hsz-2020-0169 2020

  13. [14]

    Arends, B. K. O., McCormick, J. M., van der Harst, P., Heus, P. & van Es, R. Barriers, facilitators and strategies for the implementation of artificial intelligence-based electrocardiogram interpretation: A mixed-methods study.Eur. J. Clin. Investig.55, e14387, DOI: 10.1111/eci.14387 (2025)

    work page doi:10.1111/eci.14387 2025

  14. [15]

    Swartout, W. R. & Moore, J. D. Explanation in second generation expert systems. In David, J.-M., Krivine, J.-P. & Simmons, R. (eds.)Second Generation Expert Systems, 543–585, DOI: 10.1007/978-3-642-77927-5_24 (Springer Berlin Heidelberg, Berlin/Heidelberg, Germany, 1993)

    work page doi:10.1007/978-3-642-77927-5_24 1993

  15. [16]

    & Rahwan, T

    Makovi, K., Sargsyan, A., Li, W., Bonnefon, J. & Rahwan, T. Trust within human-machine collectives depends on the perceived consensus about cooperative norms.Nat. Commun.14, DOI: 10.1038/s41467-023-38592-5 (2023)

    work page doi:10.1038/s41467-023-38592-5 2023

  16. [17]

    Samek, W., Montavon, G., Lapuschkin, S., Anders, C. J. & Müller, K.-R. Explaining deep neural networks and beyond: A review of methods and applications.Proc. IEEE109, 247–278, DOI: 10.1109/JPROC.2021.3060483 (2021)

    arXiv 2021

  17. [18]

    & Samek, W.Explainable AI Methods - A Brief Overview, 13–38 (Springer International Publishing, Cham, 2022)

    Holzinger, A., Saranti, A., Molnar, C., Biecek, P. & Samek, W.Explainable AI Methods - A Brief Overview, 13–38 (Springer International Publishing, Cham, 2022)

    2022

  18. [19]

    F., Kors, J

    Markus, A. F., Kors, J. A. & Rijnbeek, P. R. The role of explainability in creating trustworthy artificial intelligence for health care: A comprehensive survey of the terminology, design choices, and evaluation strategies.J. Biomed. Informatics 113, 103655, DOI: 10.1016/j.jbi.2020.103655 (2021)

    arXiv 2020

  19. [20]

    N.et al.Explainable artificial intelligence models using real-world electronic health record data: a systematic scoping review.J

    Payrovnaziri, S. N.et al.Explainable artificial intelligence models using real-world electronic health record data: a systematic scoping review.J. Am. Med. Informatics Assoc.27, 1173–1185, DOI: 10.1093/jamia/ocaa053 (2020)

    work page doi:10.1093/jamia/ocaa053 2020

  20. [21]

    Fusion99, 101805, DOI: 10.1016/j.inffus.2023.101805 (2023)

    Ali, S.et al.Explainable artificial intelligence (XAI): What we know and what is left to attain trustworthy artificial intelligence.Inf. Fusion99, 101805, DOI: 10.1016/j.inffus.2023.101805 (2023)

    arXiv 2023

  21. [22]

    & Schmidt, M

    Goettling, M., Hammer, A., Malberg, H. & Schmidt, M. xecgarch: a trustworthy deep learning architecture for interpretable ecg analysis considering short-term and long-term features.Sci. Reports14, 13122, DOI: 10.1038/s41598-024-63656-x (2024)

    doi:10.1038/s41598-024-63656-x 2024

  22. [23]

    Fusion106, 102301, DOI: 10.1016/j.inffus.2024.102301 (2024)

    Longo, L.et al.Explainable artificial intelligence (XAI) 2.0: A manifesto of open challenges and interdisciplinary research directions.Inf. Fusion106, 102301, DOI: 10.1016/j.inffus.2024.102301 (2024)

    arXiv 2024

  23. [24]

    & Urquhart, N

    Guckert, M., Gumpfer, N., Hannig, J., Keller, T. & Urquhart, N. A conceptual framework for establishing trust in real world intelligent systems.Cogn. Syst. Res.68, 143–155, DOI: 10.1016/j.cogsys.2021.04.001 (2021)

    work page doi:10.1016/j.cogsys.2021.04.001 2021

  24. [25]

    M.et al.A review of evaluation approaches for explainable AI with applications in cardiology.Artif

    Salih, A. M.et al.A review of evaluation approaches for explainable AI with applications in cardiology.Artif. Intell. Rev. 57, 240, DOI: 10.1007/s10462-024-10852-w (2024)

    work page doi:10.1007/s10462-024-10852-w 2024

  25. [26]

    L.et al.Novel artificial intelligence applications in cardiology: Current landscape, limitations, and the road to real-world applications.J

    Langlais, É. L.et al.Novel artificial intelligence applications in cardiology: Current landscape, limitations, and the road to real-world applications.J. Cardiovasc. Transl. Res.16, 513–525, DOI: 10.1007/s12265-022-10260-x (2023)

    work page doi:10.1007/s12265-022-10260-x 2023

  26. [27]

    M., Abu-Elkheir, M

    Elkhawaga, G., Elzeki, O. M., Abu-Elkheir, M. & Reichert, M. Why should i trust your explanation? an evaluation approach for XAI methods applied to predictive process monitoring results.IEEE Transactions on Artif. Intell.5, 1458–1472, DOI: 10.1109/TAI.2024.3357041 (2024)

    arXiv 2024

  27. [28]

    & Hannig, J

    Gumpfer, N., Dinov, B., Sossalla, S., Guckert, M. & Hannig, J. Towards Trustworthy AI in Cardiology: A Comparative Analysis of Explainable AI Methods for Electrocardiogram Interpretation. In Finkelstein, J., Moskovitch, R. & Parimbelli, E. (eds.)22nd International Conference on Artificial Intelligence in Medicine, AIME 2024, Salt Lake City, UT, USA, July ...

    work page doi:10.1007/978-3-031-66535-6_36 2024

  28. [29]

    E.et al.ACC/AHA/HRS 2008 guidelines for device-based therapy of cardiac rhythm abnormalities.Circulation 117, e350–e408, DOI: 10.1161/CIRCUALTIONAHA.108.189742 (2008)

    Epstein, A. E.et al.ACC/AHA/HRS 2008 guidelines for device-based therapy of cardiac rhythm abnormalities.Circulation 117, e350–e408, DOI: 10.1161/CIRCUALTIONAHA.108.189742 (2008)

    work page doi:10.1161/circualtionaha.108.189742 2008

  29. [30]

    Surawicz, B., Childers, R., Deal, B. J. & Gettes, L. S. AHA/ACCF/HRS recommendations for the standardization and interpretation of the electrocardiogram - part III: Intraventricular conduction disturbances.Circulation119, e235–e240, DOI: 10.1161/CIRCULATIONAHA.108.191095 (2009)

    work page doi:10.1161/circulationaha.108.191095 2009

  30. [31]

    Wagner, G. S.et al.AHA/ACCF/HRS recommendations for the standardization and interpretation of the electrocardiogram - part VI: Acute ischemia/infarction.Circulation119, e262–e270, DOI: 10.1161/CIRCULATIONAHA.108.191098 (2009)

    work page doi:10.1161/circulationaha.108.191098 2009

  31. [32]

    Samek, W., Montavon, G., Vedaldi, A., Hansen, L. K. & Müller, K.-R.Towards explainable artificial intelligence, vol. 11700 ofLecture Notes in Computer Science, 5–22 (Springer Nature, 2019)

    2019

  32. [33]

    M., Hempel, P., Maurer, M

    Metsch, J. M., Hempel, P., Maurer, M. C., Spicher, N. & Hauschild, A.-C. Ensemble post-hoc explainable AI in multivariate time series: Identifying medical features driving disease prediction.bioRxivDOI: 10.1101/2025.02.14.638219 (2025). 15/17

    work page doi:10.1101/2025.02.14.638219 2025

  33. [34]

    InIEEE International Geoscience and Remote Sensing Symposium, IGARSS 2023, Pasadena, CA, USA, July 16-21, 2023, 5014– 5017, DOI: 10.1109/IGARSS52108.2023.10282120 (IEEE, 2023)

    Najjar, H.et al.Feature attribution methods for multivariate time-series explainability in remote sensing. InIEEE International Geoscience and Remote Sensing Symposium, IGARSS 2023, Pasadena, CA, USA, July 16-21, 2023, 5014– 5017, DOI: 10.1109/IGARSS52108.2023.10282120 (IEEE, 2023)

    arXiv 2023

  34. [35]

    & Strodthoff, N

    Wagner, P., Mehari, T., Haverkamp, W. & Strodthoff, N. Explaining deep learning for ECG analysis: Building blocks for auditing and knowledge discovery.Comput. Biol. Medicine108525, DOI: 10.1016/j.compbiomed.2024.108525 (2024)

    arXiv 2024

  35. [36]

    & Samek, W

    Strodthoff, N., Wagner, P., Schaeffter, T. & Samek, W. Deep learning for ECG analysis: Benchmarks and insights from ptb-xl.IEEE J. Biomed. Heal. Informatics25, 1519–1528, DOI: 10.1109/JBHI.2020.3022989 (2021)

    arXiv 2020

  36. [37]

    Bach, S.et al.On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation.PLoS One10, 1–46, DOI: 10.1371/journal.pone.0130140 (2015)

    doi:10.1371/journal.pone.0130140 2015

  37. [38]

    Fusion99, 101883, DOI: 10.1016/j.inffus.2023.101883 (2023)

    Gumpfer, N.et al.SIGNed explanations: Unveiling relevant features by reducing bias.Inf. Fusion99, 101883, DOI: 10.1016/j.inffus.2023.101883 (2023)

    arXiv 2023

  38. [39]

    Data7, 154, DOI: 10.1038/ s41597-020-0495-6 (2020)

    Wagner, P.et al.PTB-XL, a large publicly available electrocardiography dataset.Sci. Data7, 154, DOI: 10.1038/ s41597-020-0495-6 (2020)

    2020

  39. [40]

    R.et al.Discovering and visualizing disease-specific electrocardiogram features using deep learning.Circ

    van de Leur, R. R.et al.Discovering and visualizing disease-specific electrocardiogram features using deep learning.Circ. Arrhythmia Electrophysiol.14, e009056, DOI: 10.1161/CIRCEP.120.009056 (2021)

    work page doi:10.1161/circep.120.009056 2021

  40. [41]

    Bender, T.et al.Analysis of a deep learning model for 12-lead ecg classification reveals learned features similar to diagnostic criteria.IEEE J. Biomed. Heal. Informatics28, 1848–1859, DOI: 10.1109/JBHI.2023.3271858 (2024)

    arXiv 2023

  41. [42]

    & Gross, M

    Ancona, M., Ceolini, E., Öztireli, C. & Gross, M. Towards better understanding of gradient-based attribution methods for deep neural networks. In6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings(OpenReview.net, 2018)

    2018

  42. [43]

    & Hannig, J

    Gumpfer, N., Fischer, J., Peter, L. & Hannig, J. A systematic review on explainable AI for time series classification. zenodo.orgDOI: 10.5281/zenodo.21509750 (2026). Preprint, v.1.0

    work page doi:10.5281/zenodo.21509750 2026

  43. [44]

    & Kundaje, A

    Shrikumar, A., Greenside, P. & Kundaje, A. Learning important features through propagating activation differences. In Precup, D. & Teh, Y . W. (eds.)Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, vol. 70 ofProceedings of Machine Learning Research, 3145–3153 (PMLR, 2017)

    2017

  44. [45]

    Smilkov, D., Thorat, N., Kim, B., Viégas, F. B. & Wattenberg, M. Smoothgrad: removing noise by adding noise.CoRR abs/1706.03825, DOI: 10.48550/arXiv.1706.03825 (2017)

    Pith/arXiv arXiv doi:10.48550/arxiv.1706.03825 2017

  45. [46]

    & Yan, Q

    Sundararajan, M., Taly, A. & Yan, Q. Axiomatic attribution for deep networks. In Precup, D. & Teh, Y . W. (eds.) Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, vol. 70 ofProceedings of Machine Learning Research, 3319–3328 (PMLR, 2017)

    2017

  46. [47]

    Lundberg, S. M. & Lee, S. A unified approach to interpreting model predictions. In Guyon, I.et al.(eds.)Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, 4765–4774 (2017)

    2017

  47. [48]

    D., Sturmfels, P., Lundberg, S

    Erion, G., Janizek, J. D., Sturmfels, P., Lundberg, S. M. & Lee, S.-I. Improving performance of deep learning models with axiomatic attribution priors and expected gradients.Nat. Mach. Intell.3, 620–631, DOI: 10.1038/s42256-021-00343-w (2021)

    doi:10.1038/s42256-021-00343-w 2021

  48. [49]

    Adebayo, J., Gilmer, J., Goodfellow, I. J. & Kim, B. Local explanation methods for deep neural networks lack sensitivity to parameter values. In6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Workshop Track Proceedings(OpenReview.net, 2018)

    2018

  49. [50]

    & Zisserman, A

    Simonyan, K., Vedaldi, A. & Zisserman, A. Deep inside convolutional networks: Visualising image classification models and saliency maps. In Bengio, Y . & LeCun, Y . (eds.)2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Workshop Track Proceedings, DOI: 10.48550/arXiv.1312.6034 (2014)

    Pith/arXiv arXiv doi:10.48550/arxiv.1312.6034 2014

  50. [51]

    Achtibat, R.et al.From attribution maps to human-understandable explanations through concept relevance propagation. Nat. Mach. Intell.5, 1006–1019, DOI: 10.1038/s42256-023-00711-8 (2023)

    doi:10.1038/s42256-023-00711-8 2023

  51. [52]

    General Pitfalls of Model-Agnostic Interpretation Methods for Machine Learning Models, 39–68 (Springer International Publishing, Cham, 2022)

    Molnar, C.et al. General Pitfalls of Model-Agnostic Interpretation Methods for Machine Learning Models, 39–68 (Springer International Publishing, Cham, 2022)

    2022

  52. [53]

    & Keim, D

    Schlegel, U., Arnout, H., El-Assady, M., Oelke, D. & Keim, D. A. Towards a rigorous evaluation of xai methods on time series. In2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), 4197–4201, DOI: 10.1109/ICCVW.2019.00516 (2019). 16/17

    arXiv 2019

  53. [54]

    In Wallach, H

    Chen, C.et al.This looks like that: Deep learning for interpretable image recognition. In Wallach, H. M.et al.(eds.) Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, 8928–8939 (2019)

    2019

  54. [55]

    Combi, C.et al.A manifesto on explainability for artificial intelligence in medicine.Artif. Intell. Medicine133, 102423, DOI: 10.1016/j.artmed.2022.102423 (2022)

    arXiv 2022

  55. [56]

    Houghton, A. R. The PR interval. InMaking Sense of the ECG: A Hands-On Guide, DOI: 10.1201/9780429199080-12 (CRC Press, 2019), 5 edn

    work page doi:10.1201/9780429199080-12 2019

  56. [57]

    & Vlahavas, I

    Sechidis, K., Tsoumakas, G. & Vlahavas, I. On the stratification of multi-label data.Mach. Learn. Knowl. Discov. Databases145–158, DOI: 10.1007/978-3-642-23808-6_10 (2011)

    doi:10.1007/978-3-642-23808-6_10 2011

  57. [58]

    & Kajdanowicz, T

    Szymanski, P. & Kajdanowicz, T. A network perspective on stratification of multi-label data. InFirst International Workshop on Learning with Imbalanced Domains: Theory and Applications, LIDTA@PKDD/ECML 2017, 22 September 2017, Skopje, Macedonia, vol. 74 ofProceedings of Machine Learning Research, 22–35 (PMLR, 2017)

    2017

  58. [59]

    L.et al.Physiobank, physiotoolkit, and physionet.Circulation101, e215–e220, DOI: 10.1161/01.cir.101

    Goldberger, A. L.et al.Physiobank, physiotoolkit, and physionet.Circulation101, e215–e220, DOI: 10.1161/01.cir.101. 23.e215 (2000)

    doi:10.1161/01.cir.101 2000

  59. [60]

    Kingma, D. P. & Ba, J. Adam: A method for stochastic optimization. In Bengio, Y . & LeCun, Y . (eds.)3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, DOI: 10.48550/arXiv.1412.6980 (2015)

    Pith/arXiv arXiv doi:10.48550/arxiv.1412.6980 2015

  60. [61]

    In2016 International Joint Conference on Neural Networks (IJCNN), 4368–4374, DOI: 10.1109/IJCNN.2016.7727770 (2016)

    Wang, S.et al.Training deep neural networks on imbalanced data sets. In2016 International Joint Conference on Neural Networks (IJCNN), 4368–4374, DOI: 10.1109/IJCNN.2016.7727770 (2016)

    arXiv 2016

  61. [62]

    & Guidotti, R

    Theissler, A., Spinnato, F., Schlegel, U. & Guidotti, R. Explainable AI for time series classification: A review, taxonomy and research directions.IEEE Access10, 100700–100724, DOI: 10.1109/access.2022.3207765 (2022)

    arXiv 2022

  62. [63]

    Makowski, D.et al.NeuroKit2: A python toolbox for neurophysiological signal processing.Behav. Res. Methods53, 1689–1696, DOI: 10.3758/s13428-020-01516-y (2021)

    doi:10.3758/s13428-020-01516-y 2021

  63. [64]

    Alcaraz, J. M. L. & Strodthoff, N. Diffusion-based conditional ecg generation with structured state space models.Comput. Biol. Medicine163, 107115, DOI: 10.1016/j.compbiomed.2023.107115 (2023)

    arXiv 2023

  64. [65]

    & Clifton, D

    Jarchi, D., Mahdi, A., Tarassenko, L. & Clifton, D. A. Visualisation of long-term ECG signals applied to post-intensive care patients. In2018 IEEE 15th International Conference on Wearable and Implantable Body Sensor Networks (BSN), 165–168, DOI: 10.1109/BSN.2018.8329684 (2018). Acknowledgements This work was supported by the German Federal Ministry of Re...

    arXiv 2018