Skip to main content

m45 · Evidence and regulation of AI in EP: study designs, reporting checklists, change control and monitoring

Warning

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

In 60 seconds

Most AI-in-EP studies are retrospective. A 2026 meta-analysis found that machine-learning models discriminated atrial fibrillation (AF) recurrence after ablation better than traditional risk scores, pooled over 11 studies (Koh et al. 2026); a 2025 EHRA, HRS and ESC statement found that no item of its 29-item AI reporting checklist was reported by more than 55% of the EP papers it reviewed (Svennberg et al. 2025). Moving up the evidence ladder means external validation, prospective use and, for claims of patient benefit, randomized trials, of which EP now has examples (Deisenhofer et al. 2025; Yao et al. 2021). Reporting and appraisal tools exist for each stage (Collins et al. 2024; Moons et al. 2025; Vasey et al. 2022; Cruz Rivera et al. 2020; Liu et al. 2020). Once deployed, a model meets new patients, devices and practice, so it must be monitored, and changes to it follow a plan agreed in advance (U.S. Food and Drug Administration 2024). This module derives the sample size for a sensitivity study, the net benefit of a decision curve and the CUSUM drift alarm, and the capstone lab takes an AI-ECG detector through validation, deployment, drift and a controlled update.

From the field

  • Discrimination is not enough. In the meta-analysis of AF-recurrence models, pooled AUC was 0.80 for machine-learning models against 0.58 for risk scores, with methodological quality assessed by PROBAST (Koh et al. 2026). A multicentre model that added ablation quality metrics reported an internal AUC of 0.918 and AUCs of 0.992 and 0.951 in two external hospitals (Wang et al. 2026): results to read with the appraisal questions of this module, not to adopt.
  • Reporting gaps in EP. The EHRA AI checklist (29 items) was applied to 55 EP studies published between May 2021 and May 2024; trial registration, participant details, data handling and training performance were each reported by fewer than 20% (Svennberg et al. 2025). LLMs have also been tested as reviewers against this checklist (Sluha et al. 2025).
  • Randomized evidence exists. A cluster-randomized pragmatic trial of an AI-ECG tool in primary care increased new diagnoses of low ejection fraction from 1.6% to 2.1% (Yao et al. 2021). In TAILORED-AF, a double-blind randomized trial, ablation guided by an AI algorithm in addition to pulmonary vein isolation met its primary endpoint of freedom from AF at 12 months, 88% against 70% with isolation alone (Deisenhofer et al. 2025). A blinded randomized trial with Stanford and Cedars-Sinai authors found that cardiologists changed AI first reads of ejection fraction less often than sonographers’ first reads (16.8% against 27.2% of studies substantially changed) (He et al. 2023).
  • Regulators list what they have authorized. The FDA keeps a public list of AI-enabled medical devices, notes that it is not a comprehensive resource, and states that it regulates devices, including AI-enabled ones, rather than AI as such (U.S. Food and Drug Administration 2026). Researchers at Stanford analysed how the AI devices cleared by the FDA had been evaluated and made recommendations for their evaluation (Wu et al. 2021).
  • Validation across sites without moving data (NVIDIA). A federated model trained with data from 20 institutes predicted the oxygen needs of patients with COVID-19 with an average AUC above 0.92 and generalized better than models trained at a single site (Dayan et al. 2021); NVIDIA FLARE is an open framework for such training (Roth et al. 2022).

Derivations

The evidence ladder

Rung Design What it can show Reporting or appraisal tool
0 Development and internal validation That a signal exists in these data TRIPOD+AI (Collins et al. 2024); PROBAST+AI (Moons et al. 2025)
1 External validation That it transfers to other sites, devices, patients TRIPOD+AI; PROBAST+AI
2 Prospective, non-randomized, often “silent” That it works in the live workflow DECIDE-AI (Vasey et al. 2022)
3 Randomized trial That using it changes care or outcomes SPIRIT-AI (Cruz Rivera et al. 2020), CONSORT-AI (Liu et al. 2020)
— EP-specific reporting Items particular to EP studies EHRA AI checklist (Svennberg et al. 2025)

PROBAST+AI assesses model development (16 signalling questions) and model evaluation (18) across four domains: participants and data sources, predictors, outcome and analysis (Moons et al. 2025). DECIDE-AI covers early clinical evaluation with 17 AI-specific and 10 generic items (Vasey et al. 2022). SPIRIT-AI and CONSORT-AI extend the standard trial-protocol and trial-report guidelines to interventions with an AI component (Cruz Rivera et al. 2020; Liu et al. 2020).

Key takeaways

  • Name the rung before reading the result; a retrospective AUC is not evidence of benefit.
  • Use the matching checklist when writing and when reviewing.
  • In EP, most studies are still on rungs 0 and 1 (Svennberg et al. 2025; Koh et al. 2026).

How many patients?

To estimate a sensitivity near \(Se\) with a 95% interval of half-width \(d\), a study needs about \(z^2 Se(1-Se)/d^2\) patients with the condition. If a share \(\pi\) of enrolled patients has it, the total to enrol is (Buderer 1996)

\[ n_{Se} = \frac{z^2\,Se\,(1-Se)}{d^2\,\pi}, \qquad n_{Sp} = \frac{z^2\,Sp\,(1-Sp)}{d^2\,(1-\pi)} \tag{1}\]

and the study needs the larger of the two. At \(Se = 0.9\), \(d = 0.05\) and \(\pi = 0.2\), that is 692 patients; at \(\pi = 0.05\) it is four times as many. Lab part B checks the formula by simulating studies and measuring the half-width of their Wilson intervals (m43).

Exact

Assumptions

  • Model: Equation 1 with \(z = 1.96\), the normal approximation to the binomial, independent patients and a known prevalence. The JavaScript is tested against epml.evidence.
  • Slider values are illustrative, not targets from any guideline.
  • What would change the result: several recordings per patient (the design effect of m35), a prevalence that differs from the one assumed, and sensitivities near 0 or 1, where the normal approximation is poor.

Key takeaways

  • Sensitivity is estimated only on patients with the condition: rare conditions need large studies.
  • Plan the sample size before the study, from the interval you need.

Is the model useful? Decision curves

A model can discriminate well and still not help a decision. Decision-curve analysis (Vickers and Elkin 2006) asks whether acting on the model beats acting on everyone or on no one. A threshold probability \(p_t\) expresses how a clinician weighs a false positive against a missed case: at \(p_t\) they are indifferent between acting and not. For the rule “act if the model’s probability is at least \(p_t\)”,

\[ \mathrm{NB}(p_t) = \frac{\mathrm{TP}}{n} - \frac{\mathrm{FP}}{n}\cdot\frac{p_t}{1 - p_t} \tag{2}\]

Acting on no one gives 0; acting on everyone gives \(\pi - (1-\pi)\,p_t/(1-p_t)\); a perfect model gives \(\pi\). A model is useful at the thresholds where its curve lies above both.

Key takeaways

  • Net benefit puts true and false positives on one scale, weighted by the decision’s threshold.
  • Report the curve over the range of thresholds clinicians would use.

Monitoring after deployment

A deployed model meets patients, devices and practice that differ from its validation data (m37’s dataset shift). Labels arrive late or never, but scores arrive daily. Page’s cumulative sum (Page 1954) detects a lasting upward shift in a standardized daily statistic \(x_t\):

\[ S_t = \max\bigl(0,\; S_{t-1} + x_t - \mu_0 - k\bigr), \quad S_0 = 0, \qquad \text{alarm when } S_t > h \tag{3}\]

The allowance \(k\) ignores small deviations and the threshold \(h\) sets how much evidence must accumulate. Raising \(h\) makes false alarms rarer and detection slower; for a large shift \(\delta\) the delay is close to \(h/(\delta - k)\) observations. A score alarm says that something changed, not what; finding out needs a labelled sample.

Simulation

Assumptions

  • Model: daily mean scores standardized to mean 0 and standard deviation 1 in the validation period, independent from day to day, with a step change in the mean on day 61; Equation 3 with target 0. Random numbers come from a fixed seed, so the picture is repeatable; the recursion is tested against epml.evidence.
  • Shift size, \(k\) and \(h\) are illustrative, not recommended settings.
  • What would change the result: gradual rather than step drift, weekly patterns in case mix, correlated days, and a change that alters the labels but not the scores, which a score monitor cannot see.

Key takeaways

  • Monitor what you have every day (scores, input statistics), and audit labels on a sample.
  • Choose \(k\) and \(h\) from the false-alarm rate you can afford and the delay you can accept.

Change control

AI models are updated: retrained on new data, given a new threshold, or given new inputs. For AI-enabled device software, a predetermined change control plan (PCCP), authorized with the marketing submission, lets the manufacturer make the planned changes without a new submission for each one (U.S. Food and Drug Administration 2024). The engineering habit behind it is general: write down the kinds of change you plan and the acceptance criteria before you see the results, judge each candidate on a locked test set, and treat a change the plan did not describe as needing its own review. In lab part E, neither retraining nor re-tuning the threshold met all the pre-specified criteria after a population shift; a new input did, but it was outside the plan.

Key takeaways

  • Acceptance criteria are written before the update, on interval bounds, including subgroups.
  • The locked test set must come from the population the model will meet.
  • A change outside the plan is a new review, however good its numbers.

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 AI in EP stands on the ladder. For AF recurrence after ablation, a meta-analysis pooled 11 studies of machine-learning and traditional models (Koh et al. 2026), and new multicentre models continue to appear with internal and external validation (Wang et al. 2026); these are rungs 0 and 1. Prospective, non-randomized work exists for AI-ECG screening (Noseworthy et al. 2022). Randomized trials exist for AI-ECG detection of low ejection fraction (Yao et al. 2021), for AI-guided screening for cardiomyopathy in pregnancy (Adedinsewo et al. 2024), and for AI-guided ablation in persistent AF (Deisenhofer et al. 2025). The 2025 scientific statement on AI in clinical EP reviewed reporting across AF management, sudden cardiac death and EP-lab applications and found important items under-reported (Svennberg et al. 2025).

What a randomized trial adds. It tests the whole intervention, including how clinicians act on the model’s output, against usual care, with outcomes that matter to patients. A retrospective AUC answers a narrower question: whether the model ranks patients well in data like its own.

What a reader should ask of an EP AI study:

  • Which rung? And which checklist was it reported against?
  • Which population, device and reference standard? Does the intended population match?
  • What happened after deployment? Monitoring plan, drift alarms, and the plan for updates.

Visual summary

Schematic

Development andinternal validation Externalvalidation Prospective,silent use Randomized trial:effect on care TRIPOD+AI, PROBAST+AIDECIDE-AISPIRIT-AI, CONSORT-AI Deployment:monitor scores Drift alarm:audit labels Update per plan,criteria fixed first
Figure 1: Top: the evidence ladder and its reporting tools. Bottom: after deployment, monitoring, alarms and updates that follow a plan written in advance.

Assumptions

  • Schematic: no quantitative claim. Stages follow the table of the evidence ladder and lab parts D and E.

How experts think

Before trusting an AI-in-EP result, ask:

  1. Which rung of the ladder? Internal, external, prospective or randomized; and against which checklist (Collins et al. 2024; Svennberg et al. 2025).
  2. Was the study large enough? For the interval it reports, given the prevalence (Equation 1).
  3. Does it help a decision? Net benefit over the relevant thresholds (Equation 2), and calibration (m35).
  4. What happens after deployment? Monitoring, alarms (Equation 3) and audits.
  5. How will it change? A plan with acceptance criteria fixed in advance (U.S. Food and Drug Administration 2024).

One sentence to keep: evidence is a ladder, deployment is a process, and every update is a test whose pass mark is set before the result is known.

Cheat sheet

Quantity Result
Patients for a sensitivity (Equation 1) \(n = z^2 Se(1-Se)/(d^2\pi)\); for specificity divide by \(1-\pi\)
Net benefit (Equation 2) \(\mathrm{TP}/n - (\mathrm{FP}/n)\,p_t/(1-p_t)\); treat all \(= \pi - (1-\pi)p_t/(1-p_t)\)
CUSUM (Equation 3) \(S_t = \max(0, S_{t-1} + x_t - \mu_0 - k)\); alarm at \(S_t > h\); delay \(\approx h/(\delta - k)\) for large shifts

Tools in this module

Tool For Source
TRIPOD+AI Reporting prediction-model studies Collins et al. (2024)
PROBAST+AI Quality, risk of bias and applicability Moons et al. (2025)
DECIDE-AI Early clinical evaluation Vasey et al. (2022)
SPIRIT-AI, CONSORT-AI Trial protocols and reports Cruz Rivera et al. (2020), Liu et al. (2020)
EHRA AI checklist Reporting AI research in EP Svennberg et al. (2025)
PCCP Planned changes to AI-enabled device software U.S. Food and Drug Administration (2024)

Rules of thumb

  • Report intervals, calibration and net benefit, not only AUC.
  • Monitor scores daily and labels on a sample; set alarm thresholds from the false-alarm rate.
  • Fix acceptance criteria before any update, on interval bounds and in subgroups.

Interview questions

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

Q1. A paper reports AUC 0.95 for an AF-recurrence model. What do you ask?

Model answer

Which rung: internal or external validation, and in which centres. Sample size and number of events, calibration, and net benefit at decision thresholds. How the outcome was ascertained (monitoring intensity changes recurrence rates). Whether the predictors are available at the time of the decision. Then appraise it with PROBAST+AI (Moons et al. 2025) and check reporting against TRIPOD+AI and the EHRA checklist (Collins et al. 2024; Svennberg et al. 2025).

Q2. How many patients does a study need to estimate a sensitivity of about 0.9 within ±0.05 if 20% have the condition?

Model answer

\(1.96^2 \times 0.9 \times 0.1 / (0.05^2 \times 0.2) \approx 692\) (Equation 1) (Buderer 1996), of whom about 138 have the condition. Check the specificity requirement too and enrol the larger number; inflate for several recordings per patient.

Q3. What does a decision curve show that an ROC curve does not?

Model answer

Whether acting on the model is better than acting on everyone or no one, at the threshold probabilities clinicians would actually use, with false positives weighted by the odds of the threshold (Equation 2) (Vickers and Elkin 2006). An ROC curve shows ranking only, with no costs and no prevalence.

Q4. How would you monitor a deployed AI-ECG detector when labels arrive months later?

Model answer

Daily statistics that need no labels, such as the mean score, the share of positives, and input quality and distributions, each standardized against the validation period and watched with a CUSUM (Equation 3) (Page 1954). Pair it with audits of a labelled sample, and with subgroup checks. Lab part D shows a drift that left sensitivity almost unchanged while specificity fell, mostly in one subgroup.

Q5. How do \(k\) and \(h\) change a CUSUM?

Model answer

\(k\) is the allowance: deviations smaller than \(k\) do not accumulate, so it sets the size of shift the chart is tuned to. \(h\) is the threshold: larger \(h\) means rarer false alarms and longer delays. For a large shift \(\delta\) the delay is about \(h/(\delta - k)\).

Q6. What is a predetermined change control plan, and what goes in one?

Model answer

For AI-enabled device software, a plan authorized with the marketing submission that lets the manufacturer make the planned changes without a new submission for each (U.S. Food and Drug Administration 2024). The engineering content: which changes are planned (for example retraining on new data with the same inputs), how each will be validated, the acceptance criteria fixed in advance, and how the update is rolled out and monitored. Lab part E applies the idea; it is not a regulatory template.

Q7. A retrained model fixes a subgroup but lowers overall sensitivity. Ship it?

Model answer

Only if it meets the criteria written before the update, on interval bounds, including the sensitivity bound. In lab part E neither the retrained model nor its re-tuned threshold passed; the fix that worked added an input, a change outside the plan that needs its own review.

Q8. Why might a model with excellent retrospective AUC fail a randomized trial?

Model answer

The trial tests the whole intervention: whether clinicians see, trust and act on the output, whether acting changes outcomes, and whether the trial population resembles the development data. A retrospective AUC tests none of these. EP has trials that met their endpoints (Yao et al. 2021; Deisenhofer et al. 2025), and each reports its effect for its own population and setting.

Q9. What did the EHRA AI checklist find about AI-in-EP reporting?

Model answer

Applied to 55 studies in AF management, sudden cardiac death and EP-lab applications, no item was reported by more than 55% of papers in any domain, and trial registration, participant details, data handling and training performance were reported by fewer than 20% (Svennberg et al. 2025).

Q10. How can several hospitals validate a model without pooling patient data?

Model answer

Federated learning and evaluation: each site trains or evaluates locally and shares only model updates or summary results. A federated model across 20 institutes generalized better than single-site models (Dayan et al. 2021); NVIDIA FLARE is an open framework for this (Roth et al. 2022). Governance, site differences and the reference standard still have to be checked site by site.

Q11. A drift alarm fires. What next?

Model answer

Find out what changed before acting: check input data quality and case mix, label a sample, and compare subgroups. Decide whether the model is still within its validated use; if not, restrict or pause it and follow the change plan. Record the alarm, the findings and the decision.

Q12. Which checklist goes with which study?

Model answer

TRIPOD+AI for reporting prediction-model development and validation, PROBAST+AI for appraising them (Collins et al. 2024; Moons et al. 2025); DECIDE-AI for early clinical evaluation (Vasey et al. 2022); SPIRIT-AI and CONSORT-AI for trial protocols and reports (Cruz Rivera et al. 2020; Liu et al. 2020); and the EHRA AI checklist for EP studies (Svennberg et al. 2025).

Code lab

The Part VII capstone takes the RR-interval AF detector of m37 through a model’s life: sample size, report card, decision curve, deployment, drift and a controlled update.

Part What you do Data
A Place Part VII’s studies on the evidence ladder course sources
B How many patients? (Equation 1), checked by simulation simulation
C Report card and decision curve for model v1 (Equation 2) synthetic ECGs
D Silent drift and a CUSUM alarm without labels (Equation 3) synthetic ECGs
E Change control: three candidate updates against pre-specified criteria synthetic ECGs
F Train on one database, monitor on another Challenge 2017 and PTB-XL (download)
  • Python notebook: labs/python/m45_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), checked against their published SHA-256 lists; skipped without internet access.
  • The change-control part illustrates acceptance criteria fixed in advance; its criteria are ours and are not a regulatory template.

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 condition becomes four times rarer in the study population. What happens to the patients needed for a given sensitivity interval?

Answer

Four times as many (Equation 1): the denominator holds the prevalence.

S2. At a threshold probability of 0.2, what is the net benefit of acting on everyone when prevalence is 0.2?

Answer

\(0.2 - 0.8 \times 0.25 = 0\) (Equation 2): at a threshold equal to the prevalence, acting on everyone is no better than acting on no one.

S3. \(k = 0.5\), \(h = 2\). The standardized daily values are 1.5, 1.5, 1.0. On which day does the CUSUM alarm?

Answer

\(S\) = 1.0, 2.0, 2.5 (Equation 3): the alarm is on day 3, when \(S\) first exceeds 2.

S4. Which design can show that using a model improves outcomes?

Answer

A randomized trial comparing care with and without the model, reported with CONSORT-AI (Liu et al. 2020).

S5. True or false: if a retrained model’s AUC is higher, the update can be deployed under the plan.

Answer

False. It must meet every pre-specified criterion, including sensitivity, specificity and subgroups, and be a kind of change the plan describes.

S6. Which tool assesses risk of bias in a prediction-model study?

Answer

PROBAST+AI (Moons et al. 2025).

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.

  • Decision curves. Vickers and Elkin introduced the decision curve with a worked example in prostate cancer; look at how the model’s curve compares with “treat all” and “treat none” across thresholds [Vickers and Elkin (2006); free full text]. Compare with Equation 2 and lab part C.
  • The PROBAST+AI structure. Moons and colleagues set out the two parts (development and evaluation) and four domains of the tool [Moons et al. (2025); free full text]. This module: the evidence ladder.

Further reading

  • Reporting and appraisal. Collins et al. (2024), Moons et al. (2025), Vasey et al. (2022), Cruz Rivera et al. (2020), Liu et al. (2020), and for EP Svennberg et al. (2025).
  • Methods. Buderer (1996) (sample size), Vickers and Elkin (2006) (decision curves), Page (1954) (CUSUM).
  • Trials of AI in cardiology. Yao et al. (2021), Adedinsewo et al. (2024), He et al. (2023), Deisenhofer et al. (2025).
  • Regulation. U.S. Food and Drug Administration (2024), U.S. Food and Drug Administration (2026); device evaluation at Stanford (Wu et al. 2021); m36 for device regulation in general.
  • AI competency areas. AI and society (evidence, reporting and regulation). See the AI competency map.

References

Adedinsewo, Demilade A, Andrea Carolina Morales-Lara, Bosede B Afolabi, et al. 2024. “Artificial Intelligence Guided Screening for Cardiomyopathies in an Obstetric Population: A Pragmatic Randomized Clinical Trial.” Nature Medicine 30 (10): 2897–906. https://doi.org/10.1038/s41591-024-03243-9.
Buderer, N M. 1996. “Statistical Methodology: I. Incorporating the Prevalence of Disease into the Sample Size Calculation for Sensitivity and Specificity.” Academic Emergency Medicine 3 (9): 895–900. https://doi.org/10.1111/j.1553-2712.1996.tb03538.x.
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.
Collins, Gary S., Karel G. M. Moons, Paula Dhiman, et al. 2024. “TRIPOD+AI Statement: Updated Guidance for Reporting Clinical Prediction Models That Use Regression or Machine Learning Methods.” BMJ 385: e078378. https://doi.org/10.1136/bmj-2023-078378.
Cruz Rivera, Samantha, Xiaoxuan Liu, An-Wen Chan, et al. 2020. “Guidelines for Clinical Trial Protocols for Interventions Involving Artificial Intelligence: The SPIRIT-AI Extension.” Nature Medicine 26 (9): 1351–63. https://doi.org/10.1038/s41591-020-1037-7.
Dayan, Ittai, Holger R Roth, Aoxiao Zhong, et al. 2021. “Federated Learning for Predicting Clinical Outcomes in Patients with COVID-19.” Nature Medicine 27 (10): 1735–43. https://doi.org/10.1038/s41591-021-01506-3.
Deisenhofer, Isabel, Jean-Paul Albenque, Sonia Busch, et al. 2025. “Artificial Intelligence for Individualized Treatment of Persistent Atrial Fibrillation: A Randomized Controlled Trial.” Nature Medicine 31 (4): 1286–93. https://doi.org/10.1038/s41591-025-03517-w.
He, Bryan, Alan C Kwan, Jae Hyung Cho, et al. 2023. “Blinded, Randomized Trial of Sonographer Versus AI Cardiac Function Assessment.” Nature 616 (7957): 520–24. https://doi.org/10.1038/s41586-023-05947-3.
Koh, Thomas, Stefany Jeong, Edward Dababneh, et al. 2026. “Machine-Learning Versus Traditional Scoring Methods for Prediction of Atrial Fibrillation Recurrence After Catheter Ablation: A Systematic Review and Meta-Analysis.” Current Problems in Cardiology 52 (1): 103484. https://doi.org/10.1016/j.cpcardiol.2026.103484.
Liu, Xiaoxuan, Samantha Cruz Rivera, David Moher, et al. 2020. “Reporting Guidelines for Clinical Trial Reports for Interventions Involving Artificial Intelligence: The CONSORT-AI Extension.” Nature Medicine 26 (9): 1364–74. https://doi.org/10.1038/s41591-020-1034-x.
Moons, Karel G M, Johanna A A Damen, Tabea Kaul, et al. 2025. “PROBAST+AI: An Updated Quality, Risk of Bias, and Applicability Assessment Tool for Prediction Models Using Regression or Artificial Intelligence Methods.” BMJ 388: e082505. https://doi.org/10.1136/bmj-2024-082505.
Noseworthy, Peter A, Zachi I Attia, Emma M Behnken, et al. 2022. “Artificial Intelligence-Guided Screening for Atrial Fibrillation Using Electrocardiogram During Sinus Rhythm: A Prospective Non-Randomised Interventional Trial.” Lancet 400 (10359): 1206–12. https://doi.org/10.1016/S0140-6736(22)01637-3.
Page, E. S. 1954. “Continuous Inspection Schemes.” Biometrika 41 (1–2): 100–115. https://doi.org/10.1093/biomet/41.1-2.100.
Roth, Holger R. et al. 2022. NVIDIA FLARE: Federated Learning from Simulation to Real-World. arXiv:2210.13291. https://doi.org/10.48550/arXiv.2210.13291.
Sluha, Amalia, Henrike A K Hillmann, Emma Svennberg, et al. 2025. “Reliability of Large Language Models for Reviewing Research with Artificial Intelligence in Cardiac Electrophysiology Using the European Heart Rhythm Association Artificial Intelligence Checklist.” Europace 27 (8). https://doi.org/10.1093/europace/euaf173.
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.
U.S. Food and Drug Administration. 2024. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions; Guidance for Industry and Food and Drug Administration Staff; Availability. Federal Register 89(233):96260. https://www.govinfo.gov/content/pkg/FR-2024-12-04/pdf/2024-28361.pdf.
U.S. Food and Drug Administration. 2026. Artificial Intelligence-Enabled Medical Devices. Web page with the AI-Enabled Medical Device List. https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-enabled-medical-devices.
Vasey, Baptiste, Myura Nagendran, Bruce Campbell, et al. 2022. “Reporting Guideline for the Early-Stage Clinical Evaluation of Decision Support Systems Driven by Artificial Intelligence: DECIDE-AI.” Nature Medicine 28 (5): 924–33. https://doi.org/10.1038/s41591-022-01772-9.
Vickers, Andrew J, and Elena B Elkin. 2006. “Decision Curve Analysis: A Novel Method for Evaluating Prediction Models.” Medical Decision Making 26 (6): 565–74. https://doi.org/10.1177/0272989X06295361.
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.
Wang, Kexin, Cao Zou, Xiaomeng Yin, et al. 2026. “Machine Learning Integration of Ablation Quality Metrics for Predicting Atrial Fibrillation Recurrence: A Multicenter Study.” JACC. Clinical Electrophysiology, ahead of print. https://doi.org/10.1016/j.jacep.2026.09.007.
Wu, Eric, Kevin Wu, Roxana Daneshjou, et al. 2021. “How Medical AI Devices Are Evaluated: Limitations and Recommendations from an Analysis of FDA Approvals.” Nature Medicine 27 (4): 582–84. https://doi.org/10.1038/s41591-021-01312-x.
Yao, Xiaoxi, David R Rushlow, Jonathan W Inselman, et al. 2021. “Artificial Intelligence-Enabled Electrocardiograms for Identification of Patients with Low Ejection Fraction: A Pragmatic, Randomized Clinical Trial.” Nature Medicine 27 (5): 815–19. https://doi.org/10.1038/s41591-021-01335-4.