Skip to main content

m46 · Trustworthy AI in EP: explanations, uncertainty and conformal prediction, fairness, out-of-distribution inputs and robustness

Warning

Education only. Not for diagnosis, treatment or any clinical decision.

In 60 seconds

A model that scores well on average can still mislead: its explanation may not reflect what it computes, its confidence may be wrong for one class or one group of patients, it may be applied to recordings unlike its training data, and small changes to the input can flip its output. This module builds a tool for each concern. Integrated gradients attribute an output to the samples of an ECG, and a randomization test checks that the attribution depends on the model (Sundararajan et al. 2017; Adebayo et al. 2018). Conformal prediction turns scores into sets with a coverage guarantee (Angelopoulos and Bates 2021). Group metrics show why equal error rates and equal predictive values cannot both hold when prevalences differ (Chouldechova 2017), as AI-ECG studies across race, sex and age illustrate (Noseworthy et al. 2020; Kaur et al. 2023). A Mahalanobis score flags unfamiliar inputs (Lee et al. 2018), and adversarial examples show how fragile an ECG classifier can be (Han et al. 2020; Goodfellow et al. 2014).

From the field

  • Explanations on the 12-lead ECG. A network trained on 992,748 ECGs from 365,009 patients reached an AUC of at least 0.960 for 32 of 38 classes in its test set, and a local-surrogate explanation method (LIME) highlighted physiologically relevant ECG segments (Hughes et al. 2021). At Stanford, interpretation analyses of an echocardiography network showed attention to the relevant cardiac structures for tasks people can check, and highlighted regions for tasks they cannot, such as predicting age and sex (Ghorbani et al. 2020).
  • Performance by race, sex and age. A network that detects low ejection fraction from the ECG, derived in a cohort that was 96.2% non-Hispanic white, kept AUCs between 0.931 and 0.961 across racial and ethnic subgroups of its test cohort; a separate network could predict racial category from the ECG, and the authors recommended reporting performance by race, ethnicity, age and sex (Noseworthy et al. 2020). A Stanford heart-failure model showed no significant difference between racial groups overall but performed worse in young Black patients, most of all young Black women; adding demographics to the model, training separate models or balancing the data did not remove the gap, and group-specific thresholds improved F1 (Kaur et al. 2023).
  • Attacks that experts cannot see. Smoothed adversarial perturbations fooled a single-lead arrhythmia network while remaining invisible to expert review; the authors advise care where there is an incentive to alter ECGs (Han et al. 2020).
  • Reporting. The EHRA, HRS and ESC statement found key items such as participant details and data handling under-reported in AI-in-EP studies (Svennberg et al. 2025) (m45).

Derivations

Explanations: integrated gradients

For a model output \(F(\mathbf x)\), an input \(\mathbf x\) (for example the samples of a beat) and a baseline \(\mathbf x'\) (a flat line), integrated gradients attribute to sample \(i\) (Sundararajan et al. 2017)

\[ \mathrm{IG}_i(\mathbf x) = (x_i - x'_i)\int_0^1 \frac{\partial F\bigl(\mathbf x' + \alpha(\mathbf x - \mathbf x')\bigr)}{\partial x_i}\,d\alpha, \qquad \sum_i \mathrm{IG}_i(\mathbf x) = F(\mathbf x) - F(\mathbf x') \tag{1}\]

The second identity, completeness, is the fundamental theorem of calculus along the path; checking it tells you whether the integral was approximated with enough steps. An explanation can still mislead. The model-randomization test (Adebayo et al. 2018) recomputes it for the same architecture with random weights: if little changes, it describes the input more than the model. In lab part A the attributions of a wide-QRS classifier were concentrated near the QRS complex more than those of a randomized copy (53% against 42% of the total within 60 ms of the R peak), but their ranks still correlated (0.56), because the factor \(x_i - x'_i\) carries the shape of the beat into any model’s explanation.

Key takeaways

  • Attributions should add up to the change in output; check completeness.
  • Run the randomization test before believing an explanation.
  • An explanation is a hypothesis about the model, not a reason the patient has a condition.

Uncertainty: conformal prediction

Split conformal prediction (Angelopoulos and Bates 2021) needs only a calibration set of \(n\) labelled examples that come from the same distribution as the test data. With scores \(s_i = 1 - \hat p_{y_i}(\mathbf x_i)\) (one minus the probability given to the true class) and error rate \(\alpha\),

\[ \hat q = s_{(\lceil (n+1)(1-\alpha)\rceil)}, \qquad C(\mathbf x) = \{k : 1 - \hat p_k(\mathbf x) \le \hat q\}, \qquad \Pr\bigl(y \in C(\mathbf x)\bigr) \ge 1 - \alpha \tag{2}\]

where \(s_{(j)}\) is the \(j\)-th smallest score. The guarantee is marginal: averaged over examples, not within each class or group. In lab part B a three-class rhythm model (sinus rhythm, AF, frequent early beats) had 0.936 coverage overall at \(\alpha = 0.1\) but 0.836 for early beats; one threshold per class gave 0.88 to 0.93 in each class, with sets of 1.21 classes on average instead of 1.01.

Simulation

Assumptions

  • Model: synthetic class probabilities, softmax of Gaussian logits in which each class’s own logit is raised by 3, 2 or 1 (illustrative separations); 600 calibration and 3,000 test recordings from a fixed seed. Sets follow Equation 2, with one threshold or one per class; the JavaScript is tested against epml.trust.
  • Calibration and test data are exchangeable here by construction.
  • What would change the result: a calibration set from another population or device (lab part F), class mixes that differ between calibration and use, and smaller calibration sets, whose coverage varies more.

Key takeaways

  • Conformal sets give a coverage guarantee if calibration and use data are exchangeable.
  • The guarantee is marginal; check coverage per class and per group.
  • Large or empty sets are signals: ambiguity, or an input unlike the calibration data.

Fairness: what can and cannot be equal

Write \(\pi\) for prevalence in a group, TPR and FPR for the true- and false-positive rates, and PPV for the positive predictive value. From the definitions (Chouldechova 2017),

\[ \mathrm{FPR} = \frac{\pi}{1-\pi}\cdot\frac{1-\mathrm{PPV}}{\mathrm{PPV}}\cdot\mathrm{TPR} \tag{3}\]

so if two groups have different prevalences, a model cannot have equal PPV, equal TPR and equal FPR in both. In lab part C an AF detector met two synthetic groups with 5% and 22% AF and different shares of early beats: TPR was 1.00 and 0.91, FPR 0.02 and 0.15, PPV 0.75 and 0.63. Which inequality matters depends on the decision, so report all of them by group, with intervals, and look for causes (here, early beats).

Exact

Assumptions

  • Model: PPV from prevalence, TPR and FPR (Bayes’ rule, m35) and Equation 3; one detector with the same TPR and FPR in both groups. The JavaScript is tested against epml.trust.
  • Slider values are illustrative.
  • What would change the result: error rates that themselves differ between groups (lab part C), and group-specific thresholds, which trade one inequality for another (Kaur et al. 2023).

Key takeaways

  • With different prevalences, equal error rates and equal predictive values are incompatible.
  • Report TPR, FPR and PPV per group with intervals; subgroup samples are often small.
  • Fixing a gap may need data or features, not only thresholds (Kaur et al. 2023).

Unfamiliar inputs

A model’s outputs mean little for inputs unlike its training data: a new device, a new lead, heavy noise. With features \(\mathbf f(\mathbf x)\) whose training mean is \(\boldsymbol\mu\) and covariance \(\Sigma\), the Mahalanobis distance (Lee et al. 2018)

\[ d^2(\mathbf x) = \bigl(\mathbf f(\mathbf x) - \boldsymbol\mu\bigr)^{\top}\Sigma^{-1}\bigl(\mathbf f(\mathbf x) - \boldsymbol\mu\bigr) \;\sim\; \chi^2_k \ \text{ for Gaussian features} \tag{4}\]

flags unusual inputs, and the chi-square quantile sets a false-alarm rate on familiar data. In lab part D it flagged 68% of very noisy recordings and 2% of familiar ones; accuracy on the noisy recordings was 0.53 overall and 0.93 on those not flagged. Abstaining and asking for a better recording is often the right output.

Key takeaways

  • Score how familiar an input is, and abstain when it is not.
  • Choose the threshold from the false-alarm rate on familiar data.
  • Test the score on the shifts you expect: devices, leads, noise, populations.

Robustness to small changes

For a loss \(L\), the fast gradient sign method (Goodfellow et al. 2014) moves every sample by at most \(\varepsilon\) in the direction that increases the loss fastest. For a linear score \(z = \mathbf w^{\top}\mathbf x + b\) this is exact:

\[ \mathbf x_{\text{adv}} = \mathbf x + \varepsilon\,\operatorname{sign}\bigl(\nabla_{\mathbf x} L\bigr), \qquad \lvert\Delta z\rvert \le \varepsilon\lVert\mathbf w\rVert_1, \qquad \varepsilon^\ast = \frac{\lvert z\rvert}{\lVert\mathbf w\rVert_1} \tag{5}\]

where \(\varepsilon^\ast\) is the smallest such change that reaches the decision boundary. In lab part E a linear model on raw beats needed a median of 35 µV to flip its decision, less than the 40 µV of noise in the synthetic recordings; a smoothed perturbation needed 72 µV. Smoothed attacks on a real ECG network were invisible to experts (Han et al. 2020).

Key takeaways

  • Small, structured changes can flip a classifier; noise robustness does not imply attack robustness.
  • Where there is an incentive to alter ECGs, protect the data path, not only the model (Han et al. 2020).

Clinical correlate

NotePending clinical review

This section describes published findings only. It is not guidance for patient care, and it has not yet been reviewed by an EP physician.

Where these questions arise in EP. AI-ECG models are applied to populations that differ from their development cohorts in race, sex and age; published analyses found both preserved performance (Noseworthy et al. 2020) and gaps concentrated in specific subgroups (Kaur et al. 2023). ECG interpretation tools show explanations to clinicians (Hughes et al. 2021). ECGs are acquired by more and more devices and passed through more software, which creates both unfamiliar inputs and opportunities for tampering (Han et al. 2020).

What the evidence is. The studies here are retrospective analyses of model behaviour. They show how to measure explanations, subgroup performance and robustness; none tests whether showing explanations or uncertainty to clinicians changes their decisions or patient outcomes, which would need the prospective and randomized designs of m45.

What a reader should ask:

  • Subgroups. Is performance reported by sex, age, race and ethnicity, and device, with intervals?
  • Uncertainty. Does the tool say when it is unsure or when the input is unfamiliar?
  • Explanations. Were they tested, for example with a randomization check, before being shown to users?

Visual summary

Schematic

ECG input(any device) Model Prediction setor abstain Unfamiliar? (d²) Attack test (ε*) Attributions +randomization test Conformal sets Metrics by groupwith intervals
Figure 1: Five checks around one model: before it (is the input familiar? could it have been altered?), inside it (what drives the output?) and after it (how sure is it, and is it equally good for every group?).

Assumptions

How experts think

Before trusting an AI-ECG or EP model, ask:

  1. What does it say when unsure? Sets, abstention, an unfamiliarity score (Equation 2 and Equation 4).
  2. For whom does it work? TPR, FPR and PPV by group with intervals, and the reason for any gap (Noseworthy et al. 2020; Kaur et al. 2023).
  3. What is the explanation for? Debugging and hypotheses, checked by a randomization test (Adebayo et al. 2018), not as proof.
  4. What inputs will it really see? Devices, leads, noise, and whether anyone could alter them (Han et al. 2020).
  5. Were these checks prespecified? Subgroups and thresholds chosen before testing (m45).

One sentence to keep: a trustworthy model knows when it does not know, works for each group it serves, and does not change its mind for a change nobody can see.

Cheat sheet

Quantity Result
Integrated gradients (Equation 1) \((x_i - x'_i)\int_0^1 \partial_i F(\mathbf x' + \alpha(\mathbf x - \mathbf x'))\,d\alpha\); sum \(= F(\mathbf x) - F(\mathbf x')\)
Conformal sets (Equation 2) \(\hat q\) = rank \(\lceil (n+1)(1-\alpha)\rceil\) score; \(C = \{k : 1 - \hat p_k \le \hat q\}\); coverage \(\ge 1-\alpha\)
FPR identity (Equation 3) \(\mathrm{FPR} = \frac{\pi}{1-\pi}\frac{1-\mathrm{PPV}}{\mathrm{PPV}}\mathrm{TPR}\)
Mahalanobis score (Equation 4) \(d^2 = (\mathbf f - \boldsymbol\mu)^\top\Sigma^{-1}(\mathbf f - \boldsymbol\mu) \sim \chi^2_k\)
Attack on a linear score (Equation 5) \(\lvert\Delta z\rvert \le \varepsilon\lVert\mathbf w\rVert_1\); \(\varepsilon^\ast = \lvert z\rvert/\lVert\mathbf w\rVert_1\)

Studies in this module

Study What was examined Design
Hughes et al. (2021) 12-lead ECG network with LIME explanations Cross-sectional, single centre
Ghorbani et al. (2020) Interpretation of an echocardiography network (Stanford) Retrospective
Noseworthy et al. (2020) Low-ejection-fraction AI-ECG across race and ethnicity Retrospective
Kaur et al. (2023) Heart-failure AI-ECG across race, sex and age (Stanford) Retrospective
Han et al. (2020) Smoothed adversarial examples for an ECG network Method study

Rules of thumb

  • Check completeness and run a randomization test before showing an explanation.
  • Calibrate conformal sets on data like the deployment data; check coverage per class and group.
  • Report group metrics with intervals; plan subgroup sample sizes (m45).

Interview questions

Try each question before you open its model answer. In a real interview, say which assumptions you are making.

Q1. What is integrated gradients, and what does completeness tell you?

Model answer

It attributes a model’s output to its inputs by integrating the gradient along the straight path from a baseline to the input, times the input’s difference from the baseline (Equation 1) (Sundararajan et al. 2017). Completeness says the attributions sum to \(F(\mathbf x) - F(\mathbf x')\); if they do not, the integral needs more steps.

Q2. How would you check that a saliency map for an ECG model is meaningful?

Model answer

Run the model-randomization test: recompute it with random weights and compare (Adebayo et al. 2018). Check that it is stable to tiny input changes, that removing highly attributed samples changes the output more than removing random ones, and that it highlights features experts can verify for tasks they understand (Hughes et al. 2021; Ghorbani et al. 2020). Treat it as a hypothesis about the model.

Q3. What does a conformal prediction set guarantee, and what not?

Model answer

That the set contains the true label with probability at least \(1-\alpha\), averaged over examples, if calibration and test data are exchangeable (Equation 2) (Angelopoulos and Bates 2021). Not coverage within each class or group, and not under dataset shift. Class-conditional calibration restores per-class coverage at the cost of larger sets.

Q4. Why can’t a model have equal PPV and equal error rates in two groups?

Model answer

Because \(\mathrm{FPR} = \frac{\pi}{1-\pi}\frac{1-\mathrm{PPV}}{\mathrm{PPV}}\mathrm{TPR}\) (Equation 3) (Chouldechova 2017): with equal PPV and TPR, different prevalences force different FPRs. Decide which quantity matters for the decision, and report all of them.

Q5. A model’s AUC is the same across racial groups. Is it fair?

Model answer

Not necessarily. AUC is threshold-free; error rates and PPV at the operating threshold can still differ, and gaps can hide in intersections such as age by sex by race. A Stanford heart-failure model had no significant difference between racial groups overall but performed worse in young Black patients (Kaur et al. 2023); another model kept similar AUCs across groups (Noseworthy et al. 2020). Check intersections, thresholds and intervals.

Q6. How do you detect that a recording is outside what the model was built for?

Model answer

Score its features against the training distribution, for example with a Mahalanobis distance as in Equation 4, set the threshold from the false-alarm rate on familiar data, and abstain above it. Validate the score on expected shifts: new devices, leads, noise. Empty or large conformal sets are a second signal.

Q7. What is an adversarial example, and does it matter for ECG models?

Model answer

An input changed slightly so the model’s output changes. For a linear score a max-norm change \(\varepsilon\) moves the score by up to \(\varepsilon\lVert\mathbf w\rVert_1\) (Equation 5) (Goodfellow et al. 2014). Smoothed perturbations fooled an ECG network while invisible to experts (Han et al. 2020); it matters where someone has an incentive to alter ECGs, and the defence includes securing the data path.

Q8. A subgroup has 30 patients. What can you say about its sensitivity?

Model answer

Little, unless the interval is narrow: compute a Wilson interval (m43) and plan subgroup sizes in advance (m45). In lab part F each PTB-XL subgroup holds only a few dozen recordings, so its intervals are wide.

Q9. How would you present uncertainty to a clinician using an AI-ECG tool?

Model answer

As something actionable: a set of possible rhythms at a stated error rate, a clear “unsure, repeat the recording” when the input is unfamiliar, and the tool’s validated population. Whether this improves decisions is an empirical question for a prospective study (m45), not something to assume.

Q10. Group-specific thresholds reduce a performance gap. What are the risks?

Model answer

They trade one inequality for another (Equation 3), need reliable group membership, and can be overfitted to small subgroups. They improved F1 in one study (Kaur et al. 2023); any such choice should be prespecified and tested on data not used to choose it.

Q11. Why do integrated gradients from a random model still look like the beat?

Model answer

Because each attribution is multiplied by \(x_i - x'_i\) (Equation 1), so large-amplitude samples get large attributions whatever the model. That is why the randomization test compares trained and random models rather than judging one map by eye (Adebayo et al. 2018).

Q12. Conformal sets calibrated on one database are used on another. What happens?

Model answer

The guarantee no longer holds, because the data are not exchangeable; coverage can fall below or rise above the target. Recalibrate on data from the deployment setting, and monitor coverage on labelled samples (m45).

Code lab

The lab builds each trust check on synthetic ECGs and then applies the subgroup and conformal checks to real recordings.

Part What you do Data
A Integrated gradients and the randomization check (Equation 1) synthetic beats
B Conformal sets for three rhythm classes (Equation 2) synthetic ECGs
C Error rates and predictive values in two groups (Equation 3) synthetic ECGs
D An unfamiliarity score and abstention (Equation 4) synthetic ECGs
E How small an attack flips a decision (Equation 5) synthetic beats
F Subgroups and conformal coverage Challenge 2017 and PTB-XL (download)
G Optional: integrated gradients with PyTorch autograd synthetic beats
  • Python notebook: labs/python/m46_lab.ipynb. Run it from the course repository after pip install -r requirements.txt, or open it in Google Colab.
  • Data for part F: the PhysioNet/CinC Challenge 2017 training set (Clifford et al. 2017; Clifford et al. 2018) (Open Data Commons Attribution v1.0) and PTB-XL 1.0.3 (Wagner et al. 2020) (Creative Commons Attribution 4.0), with PTB-XL’s age and sex fields as the database codes them; checked against the published SHA-256 lists and skipped without internet access.
  • Part G runs only if PyTorch is installed.

Each part ends with checks: lines that pass with the code as given. If you change something and a check fails, that is information, not an error to hide.

Self-check

Short checks you can do in your head. Open each answer after you have tried it.

S1. A model’s output goes from 0.1 at the baseline to 0.8 at the input. What do the integrated gradients add up to?

Answer

0.7 (Equation 1).

S2. 99 calibration scores, \(\alpha = 0.1\). Which ranked score is the threshold?

Answer

The \(\lceil 100 \times 0.9\rceil = 90\)th smallest (Equation 2).

S3. Prevalence 0.2, TPR 0.9, PPV 0.75. What is the FPR?

Answer

\(0.25 \times \tfrac{1}{3} \times 0.9 = 0.075\) (Equation 3).

S4. A linear score is \(z = 0.5\) with \(\lVert\mathbf w\rVert_1 = 25\) per mV. What is the smallest max-norm change that can flip the decision?

Answer

\(0.5 / 25 = 0.02\) mV, or 20 µV (Equation 5).

S5. With two Gaussian features, what share of familiar inputs exceeds the 95% chi-square threshold?

Answer

About 5% (Equation 4): the threshold is set by that false-alarm rate.

S6. True or false: coverage of 0.90 overall means each rhythm class is covered 90% of the time.

Answer

False. Coverage is marginal; in lab part B one class was covered 84% of the time while the overall figure was 94%.

Literature figures

These figures are worth looking at alongside this module. We link to them instead of copying them: some licenses do not allow reuse here, and others we could not confirm. Each entry says what to look for and where this module explains it.

  • Explanations on ECGs. Hughes and colleagues use LIME to visualize the ECG segments that contributed to the network’s diagnoses [Hughes et al. (2021); free full text]. Compare with lab part A.
  • Adversarial ECGs. Han and colleagues construct smoothed adversarial examples for single-lead ECGs; compare the perturbed tracings with the originals [Han et al. (2020); free full text]. This module: robustness.
  • Performance by subgroup. Noseworthy and colleagues report AUC by racial and ethnic subgroup [Noseworthy et al. (2020); free full text]. This module: fairness.

Further reading

  • Explanations. Sundararajan et al. (2017), Adebayo et al. (2018); in cardiology Hughes et al. (2021), Ghorbani et al. (2020).
  • Uncertainty. Angelopoulos and Bates (2021); out-of-distribution detection Lee et al. (2018).
  • Fairness. Chouldechova (2017); in AI-ECG Noseworthy et al. (2020), Kaur et al. (2023), and Hsieh et al. (2026) (m43).
  • Robustness. Goodfellow et al. (2014), Han et al. (2020).
  • Next module. m47: the data and deployment engineering behind a model you can trust.
  • AI competency areas. AI and society (fairness and robustness); planning and game playing (adversarial min-max). See the AI competency map.

References

Adebayo, Julius et al. 2018. Sanity Checks for Saliency Maps. arXiv:1810.03292. https://doi.org/10.48550/arXiv.1810.03292.
Angelopoulos, Anastasios N., and Stephen Bates. 2021. A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification. arXiv:2107.07511. https://doi.org/10.48550/arXiv.2107.07511.
Chouldechova, Alexandra. 2017. Fair Prediction with Disparate Impact: A Study of Bias in Recidivism Prediction Instruments. arXiv:1703.00056. https://doi.org/10.48550/arXiv.1703.00056.
Clifford, Gari D. et al. 2017. AF Classification from a Short Single Lead ECG Recording: The PhysioNet/Computing in Cardiology Challenge 2017 (Version 1.0.0). PhysioNet. https://physionet.org/content/challenge-2017/1.0.0/.
Clifford, Gari D, Chengyu Liu, Benjamin Moody, et al. 2018. “AF Classification from a Short Single Lead ECG Recording: The PhysioNet/Computing in Cardiology Challenge 2017.” Computing in Cardiology 44. https://doi.org/10.22489/CinC.2017.065-469.
Ghorbani, Amirata, David Ouyang, Abubakar Abid, et al. 2020. “Deep Learning Interpretation of Echocardiograms.” NPJ Digital Medicine 3: 10. https://doi.org/10.1038/s41746-019-0216-8.
Goodfellow, Ian J., Jonathon Shlens, and Christian Szegedy. 2014. Explaining and Harnessing Adversarial Examples. arXiv:1412.6572. https://doi.org/10.48550/arXiv.1412.6572.
Han, Xintian, Yuxuan Hu, Luca Foschini, et al. 2020. “Deep Learning Models for Electrocardiograms Are Susceptible to Adversarial Attack.” Nature Medicine 26 (3): 360–63. https://doi.org/10.1038/s41591-020-0791-x.
Hsieh, P Nelson, Parth Agrawal, Aman Alok, et al. 2026. “Artificial Intelligence Electrocardiography for Left Ventricular Systolic Dysfunction Demonstrates Preserved Performance Across Demographic Training Imbalances.” European Heart Journal. Digital Health 7 (5): ztag080. https://doi.org/10.1093/ehjdh/ztag080.
Hughes, J Weston, Jeffrey E Olgin, Robert Avram, et al. 2021. “Performance of a Convolutional Neural Network and Explainability Technique for 12-Lead Electrocardiogram Interpretation.” JAMA Cardiology 6 (11): 1285–95. https://doi.org/10.1001/jamacardio.2021.2746.
Kaur, Dhamanpreet, J Weston Hughes, Albert J Rogers, et al. 2023. “Race, Sex, and Age Disparities in the Performance of ECG Deep Learning Models Predicting Heart Failure.” Circulation. Heart Failure 17 (1): e010879. https://doi.org/10.1161/CIRCHEARTFAILURE.123.010879.
Lee, Kimin et al. 2018. A Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversarial Attacks. arXiv:1807.03888. https://doi.org/10.48550/arXiv.1807.03888.
Noseworthy, Peter A, Zachi I Attia, LaPrincess C Brewer, et al. 2020. “Assessing and Mitigating Bias in Medical Artificial Intelligence: The Effects of Race and Ethnicity on a Deep Learning Model for ECG Analysis.” Circulation. Arrhythmia and Electrophysiology 13 (3): e007988. https://doi.org/10.1161/CIRCEP.119.007988.
Sundararajan, Mukund, Ankur Taly, and Qiqi Yan. 2017. Axiomatic Attribution for Deep Networks. arXiv:1703.01365. https://doi.org/10.48550/arXiv.1703.01365.
Svennberg, Emma, Janet K Han, Enrico G Caiani, et al. 2025. “State of the Art of Artificial Intelligence in Clinical Electrophysiology in 2025: A Scientific Statement of the European Heart Rhythm Association (EHRA) of the ESC, the Heart Rhythm Society (HRS), and the ESC Working Group on e-Cardiology.” Europace 27 (5). https://doi.org/10.1093/europace/euaf071.
Wagner, Patrick, Nils Strodthoff, Ralf-Dieter Bousseljot, et al. 2020. “PTB-XL, a Large Publicly Available Electrocardiography Dataset.” Scientific Data 7 (1): 154. https://doi.org/10.1038/s41597-020-0495-6.