Skip to main content

m47 · Data and deployment engineering for AI in EP: data formats, labelling, de-identification, versioning, quantization and model cards

Warning

Education only. Not for diagnosis, treatment or any clinical decision. Nothing here is legal advice.

In 60 seconds

Most failures of AI in practice start outside the model. A file header with the wrong gain changes every amplitude; labels from experts who disagree cap what any model can score; a released table can identify the people in it; a dataset changes without its name changing; a model shrunk to fit a wearable is a different model. This module turns each of these into something you can compute and check: the header conversion \(v = (d - b)/g\) (PhysioNet 2026), the accuracy of a majority of labellers, the re-identification risk \(1/f\) of a record that shares its values with \(f - 1\) others (El Emam and Dankar 2008), a checksum manifest that fingerprints a dataset version (Wilkinson et al. 2016), the step size and noise of integer quantization (Jacob et al. 2017; Wu et al. 2020), and a model card that records all of it (Mitchell et al. 2018; Gebru et al. 2018).

From the field

  • Labels are measurements too. In the PhysioNet/CinC Challenge 2017, disagreement between experts was high enough that the organizers relabelled contentious recordings during the competition (Clifford et al. 2018). PTB-XL records were annotated by up to two cardiologists, and the database ships diagnosis likelihoods and manually annotated signal properties with the labels (Wagner et al. 2020); a well-known single-lead study used a test set annotated by a consensus committee of cardiologists (Hannun et al. 2019).
  • The ECG itself can identify a person. A deep network trained on about 40,000 patients identified individuals from their ECGs with 94.56% accuracy; accuracy varied by heart condition, for example 78.54% for patients with premature ventricular contractions (Ghazarian et al. 2021).
  • De-identification of text is a model to validate. A Stanford pipeline that finds and replaces protected health information in radiology reports reached F1 scores of 97.9 on reports from a known institution and 99.6 from a new one (Chambon et al. 2023).
  • Small devices. A 2,680-parameter arrhythmia network was pruned, quantized and run on a microcontroller with 64 KB of memory; its accuracy on MIT-BIH was 98.07% overall but 80.73% when evaluated patient-independently at record level (Huang et al. 2026), a reminder that the split decides the number (m35). Distilling a larger network into a smaller one raised the F-score after quantization from 0.779 to 0.828 in an AF-detection study on a low-power chip (Faraone et al. 2021).
  • Integer arithmetic (NVIDIA). An NVIDIA study reviewed the choices behind integer quantization and reported an 8-bit workflow that kept accuracy within 1% of the floating-point baseline on all the vision, speech and language networks it studied (Wu et al. 2020).
  • Documentation (Stanford). A Stanford-led analysis of 32,111 model cards on Hugging Face found that sections on limitations, evaluation and environmental impact were filled in least often, and training details most consistently (Liang et al. 2024).

Derivations

Data formats: units live in the header

A WFDB record stores integers \(d\) in analog-to-digital converter units (adu); its header gives the gain \(g\) in adu per physical unit and the baseline \(b\), the stored value that means zero (PhysioNet 2026). The physical value is

\[ v = \frac{d - b}{g}, \qquad \lvert v_{\text{stored}} - v \rvert \le \frac{1}{2g} \tag{1}\]

where the bound is the rounding error when a physical signal is written as integers. The format documentation says that a missing gain means the amplitude is uncalibrated and that 200 adu per physical unit may be assumed, and that missing units may be taken as millivolts (WFDB Python contributors 2022): a file with a wrong or missing header still opens. In lab part A a header claiming gain 1000 instead of 200 shrank every amplitude five-fold, and one claiming 500 instead of 250 samples per second doubled every heart rate. An audit that only learned the 1st to 99th percentiles of R amplitude and heart rate from training data flagged all 30 recordings in each case and 0–3% of correct ones.

Key takeaways

  • Read the header before the signal: rate, gain, baseline and units decide every later number.
  • Audit new data against ranges learned from training data; a flag means “look”, not “reject”.
  • A model whose features are ratios (for example RR variability) can hide a wrong sampling rate.

Labels: what a majority vote can fix

If each of \(m\) labellers (odd \(m\)) is right with probability \(p\), independently, the majority is right with probability

\[ P_{\text{maj}}(m, p) = \sum_{j=(m+1)/2}^{m} \binom{m}{j}\, p^{j} (1 - p)^{m-j} \tag{2}\]

which grows toward 1 with \(m\) when \(p > 0.5\). Real labellers share hard cases, and a vote cannot fix a shared mistake. In lab part C, labellers who were right 79% of the time gave a seven-labeller majority of 0.96 when independent, but 0.91 when 20% of recordings were hard for everyone. Agreement is measured with Cohen’s kappa (m35), and noisy reference labels cap the accuracy you can measure (m35): a model that is right 95% of the time scored 0.76 against one labeller in the lab.

Key takeaways

  • Votes fix independent errors, not shared ones; adjudicate hard cases.
  • Report how labels were made: who, how many, agreement, and what was done with disagreement.
  • Keep label versions; relabelling changes every reported metric (Clifford et al. 2018).

De-identification and re-identification risk

Fields such as age, sex, region and procedure date are quasi-identifiers: harmless alone, identifying together. Group the records that share all released quasi-identifier values; record \(i\) is in a group of size \(f_i\). Someone who knows a person’s values, and that the person is in the data, picks the right record with probability \(1/f_i\). Over a table of \(n\) records in \(J\) groups,

\[ r_i = \frac{1}{f_i}, \qquad R_{\max} = \frac{1}{\min_i f_i} = \frac{1}{k}, \qquad \bar R = \frac{1}{n}\sum_{i=1}^{n} \frac{1}{f_i} = \frac{J}{n} \tag{3}\]

where \(k\) is the smallest group size: the table is \(k\)-anonymous (El Emam and Dankar 2008). In a synthetic registry of 5,000 procedures (lab part D), 99.6% of records were unique with age, sex, device, region and procedure day released. Reaching \(k = 5\) took dropping the date and the region, ten-year age groups, and grouping the youngest and oldest ages. In the United States, HIPAA allows de-identification by expert determination of very small risk or by the Safe Harbor method, which removes a list of identifier types (18 types) and requires no actual knowledge that the rest identifies someone (U.S. Department of Health and Human Services, Office for Civil Rights 2012; U.S. Department of Health and Human Services 2010). HHS guidance cites estimates that date of birth, sex and five-digit ZIP code together are unique for over half of US residents (U.S. Department of Health and Human Services, Office for Civil Rights 2012).

Simulation

Assumptions

  • Model: a synthetic registry of 2,000 procedures from a fixed seed: age from a normal distribution (mean 65, standard deviation 12 years, limited to 18–95), a sex code, 30 region codes used unevenly (weights 1, 1/2, …, 1/30), a procedure day of one year and three device models in shares 60%, 30% and 10%. All distributions are illustrative; there are no real people.
  • Risk follows Equation 3 for an attacker who knows the person is in the registry; the JavaScript is tested against epml.deploy.
  • What would change the result: a larger registry (more records share values), an attacker who does not know who is included (lower risk), outside data that adds fields, and the waveforms themselves, which can identify people (Ghazarian et al. 2021).

Key takeaways

  • Count unique records before release; most risk sits in rare combinations and extreme values.
  • Coarsening trades detail for privacy; decide which fields the analysis really needs.
  • Table de-identification does not cover waveforms or free text; each needs its own assessment (Chambon et al. 2023).

Versions: a fingerprint for every dataset

A SHA-256 checksum changes when any bit of a file changes. A manifest lists the checksum of every file, and the checksum of the manifest identifies the whole dataset version. This course pins the checksum lists of the PhysioNet databases it downloads in data/registry.json. In lab part B flipping one bit in one signal file changed the dataset version and named the file. Persistent identifiers and rich metadata make data findable and reusable by machines as well as people (Wilkinson et al. 2016).

Key takeaways

  • Record the data version next to every result, model and model card.
  • Verify downloads against published checksums; pin the list itself.

Deployment: integer quantization

Quantization stores a real value \(r\) as an integer \(q\) with a scale \(S\) and zero point \(Z\) (Jacob et al. 2017):

\[ r = S\,(q - Z), \qquad S = \frac{r_{\max} - r_{\min}}{q_{\max} - q_{\min}}, \qquad Z = q_{\min} - \operatorname{round}\!\left(\frac{r_{\min}}{S}\right) \tag{4}\]

where \([r_{\min}, r_{\max}]\) is the chosen range (widened to include 0, so that zero is exact) and \([q_{\min}, q_{\max}]\) the integer range, for example 0 to 255 for 8 bits. With 8-bit weights and activations the memory is close to a quarter of 32-bit floating point (Jacob et al. 2017). Inside the range the rounding error is spread evenly over one step, so

\[ \sigma_e^2 = \frac{S^2}{12}, \qquad \mathrm{SQNR} = 10 \log_{10} \frac{\overline{x^2}}{S^2/12}\ \text{dB} \;\approx\; 6.02\,b + 1.76\ \text{dB for a full-scale sine} \tag{5}\]

where SQNR is the signal-to-quantization-noise ratio, \(\overline{x^2}\) the mean square of the signal and \(b\) the number of bits; each extra bit adds about 6 dB. A range wider than the signal wastes levels: in lab part E an 8-bit beat had an SQNR of 44 dB over its own range and 26 dB over a fixed ±5 mV range. For the m46 network, 8-bit weights left the AUC unchanged (0.961) at a quarter of the size; 4-bit weights changed 2.4% of decisions with one scale per tensor and 1.2% with one scale per hidden unit.

Approximate model

Assumptions

  • Model: the synthetic m37 beat (Gaussian waves, illustrative amplitudes) at 250 samples per second, quantized with Equation 4 over its own range or over a fixed ±5 mV range (an illustrative value, not a device specification). The quantizer is tested against epml.deploy.
  • The predicted line uses the uniform-noise model of Equation 5; it ignores that at very few bits the rounding error follows the signal’s shape and is no longer uniform, so measurement and prediction differ there.
  • What would change the result: noise already in the signal (which dithers the rounding), clipping when the range is too narrow, and the analog front end (m14).

Key takeaways

  • Each bit is worth about 6 dB; choosing the range well is worth as much as several bits.
  • One scale per output channel helps when weights in different rows have different ranges.
  • A quantized model is a new model: test it on the same data and subgroups as the original (m45, m46).

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 issues arise in EP. ECGs and electrograms move from recording systems into hospital archives and research databases, often through exports that carry their own headers and units. AI-ECG and EP datasets are labelled by clinicians who disagree on hard cases (Clifford et al. 2018; Wagner et al. 2020). Multicentre projects share data or train where the data live (federated learning, m45). Detection algorithms increasingly run on wearables and implants with little memory (Huang et al. 2026; Faraone et al. 2021).

What the evidence is. The studies here describe datasets, methods and engineering results. None shows how data-pipeline errors, label quality or quantization change patient outcomes; the examples in this module are synthetic.

What a reader should ask:

  • Data. Which database version, sampling rate, gain and units? How were labels made and adjudicated?
  • Privacy. How was the data de-identified, and is the waveform itself considered identifying (Ghazarian et al. 2021)?
  • Deployment. Was the model evaluated in the form that runs on the device, quantized and on the device’s own recordings?

Visual summary

Schematic

Record and export Label De-identify Version Deploy Header auditv = (d − b)/g Votes and kappaP_maj(m, p) Class sizesrisk 1/f Checksummanifest S²/12 anda model card
Figure 1: Five steps between a recording and a deployed model, each with the check this module adds.

Assumptions

How experts think

Before trusting a dataset or a deployed model, ask:

  1. What exactly is in the files? Sampling rate, gain, baseline, units, filters, and the database version (Equation 1).
  2. Where do the labels come from? Number of labellers, agreement, adjudication and relabelling, as in Equation 2 and the kappa of m35.
  3. Who could be identified, and how? Class sizes, rare values, waveforms and text (Equation 3).
  4. Can I reproduce this number? Data version, code version, model checksum.
  5. Is the deployed model the tested model? Quantization, device, input pipeline (Equation 4 and Equation 5).

One sentence to keep: every number about a model is also a number about its data pipeline; write both down.

Cheat sheet

Quantity Result
Header conversion (Equation 1) \(v = (d - b)/g\); rounding error at most \(1/(2g)\)
Majority of \(m\) labellers (Equation 2) \(\sum_{j=(m+1)/2}^{m}\binom{m}{j}p^j(1-p)^{m-j}\); for \(m = 3\): \(3p^2 - 2p^3\)
Re-identification risk (Equation 3) \(r_i = 1/f_i\); largest \(1/k\); average \(J/n\)
Affine quantization (Equation 4) \(r = S(q - Z)\), \(S = (r_{\max} - r_{\min})/(q_{\max} - q_{\min})\)
Quantization noise (Equation 5) \(\sigma_e^2 = S^2/12\); about 6 dB per bit

Sources in this module

Source What it adds Kind
Wagner et al. (2020), Clifford et al. (2018), Hannun et al. (2019) How public and published ECG datasets were labelled Dataset papers and a retrospective study
Ghazarian et al. (2021) Re-identification of patients from ECGs Method study
Chambon et al. (2023) De-identification of report text (Stanford) Retrospective method study
Huang et al. (2026), Faraone et al. (2021) Arrhythmia models on microcontrollers Method studies
Jacob et al. (2017), Wu et al. (2020) Integer quantization (Google; NVIDIA) Method studies
Mitchell et al. (2018), Gebru et al. (2018), Liang et al. (2024) Model cards, datasheets, and how cards are filled in (Stanford) Proposals and an analysis

Rules of thumb

  • Audit headers and amplitudes on every new data source.
  • Count unique records before any release; remember waveforms and free text.
  • Quantize, then retest on the same data and subgroups; record checksums in the model card.

Interview questions

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

Q1. A model trained on one database scores poorly on another. Before blaming the model, what do you check?

Model answer

The data pipeline: sampling rate, gain, baseline and units in the headers (Equation 1), filters, lead placement, and how labels were made. Compare amplitude and heart-rate distributions with the training data; a wrong header or a different filter can look like a model failure (PhysioNet 2026).

Q2. Why does a missing gain in a WFDB header matter?

Model answer

The amplitude is then uncalibrated, and a default of 200 adu per physical unit may be assumed (WFDB Python contributors 2022), so the file opens without an error but every amplitude may be wrong by a constant factor. Amplitude-based features and models change; ratio features may not, which hides the problem.

Q3. You have three labellers who are each right 80% of the time. How accurate is their majority?

Model answer

\(3p^2 - 2p^3 = 0.896\) if their errors are independent (Equation 2). If they share hard cases, less: the vote cannot fix a mistake everyone makes. Measure agreement with kappa (m35) and adjudicate disagreements.

Q4. Why can’t you measure a model’s accuracy above the accuracy of its labels?

Model answer

Against labels that are wrong at rate \(\epsilon\), a model with true accuracy \(a\) scores \(a(1 - \epsilon) + (1 - a)\epsilon\) if the errors are independent (m35). A perfect model scores \(1 - \epsilon\). Better reference labels (consensus, adjudication (Hannun et al. 2019)) raise the ceiling.

Q5. What is a quasi-identifier, and how do you measure the risk it creates?

Model answer

A field that identifies people only in combination, such as age, sex, region or a procedure date. Group records by their released values; a record in a group of \(f\) has risk \(1/f\) against someone who knows the person is in the data, the table’s largest risk is \(1/k\) and the average is groups over records (Equation 3) (El Emam and Dankar 2008).

Q6. A table has been de-identified. Can you release the ECG waveforms with it?

Model answer

Not without a separate assessment: deep networks have identified people from their ECGs with high accuracy (Ghazarian et al. 2021). Waveforms and free-text reports need their own de-identification and often a data-use agreement; text de-identification tools are models with their own error rates (Chambon et al. 2023).

Q7. How do you know two teams used the same dataset?

Model answer

Compare checksum manifests, or a single identifier computed from the manifest, not file names or dataset names. Verify downloads against the publisher’s checksum list and pin that list; record the version with every result (Wilkinson et al. 2016).

Q8. Explain scale and zero point in 8-bit quantization.

Model answer

A real value is stored as \(r = S(q - Z)\) (Equation 4) (Jacob et al. 2017): \(S\) maps one integer step to a real step, and \(Z\) is the integer that represents real zero exactly. The range is chosen from the data or a calibration set; values outside it are clipped.

Q9. How much does one extra bit improve the signal-to-quantization-noise ratio?

Model answer

About 6 dB: the step halves, so the noise variance \(S^2/12\) falls four-fold (Equation 5). For a full-scale sine the SQNR is about \(6.02b + 1.76\) dB. A range much wider than the signal costs the same as losing bits.

Q10. A model is quantized to 8 bits for a wearable. What do you test before deployment?

Model answer

The quantized model on the same test data and subgroups as the original: discrimination, calibration, decision changes at the operating threshold, and performance on recordings from the device itself. An 8-bit workflow can stay within 1% of floating-point accuracy (Wu et al. 2020), but that is a result to confirm, not assume; treat it as a model change (m45).

Q11. What goes into a model card for an AI-ECG model?

Model answer

Intended use and out-of-scope uses, training and evaluation data with versions, metrics overall and by group with intervals, limitations, and the checksum of the deployed weights (Mitchell et al. 2018). A datasheet does the same for the data (Gebru et al. 2018). Limitations and evaluation are the sections most often left empty (Liang et al. 2024).

Q12. A paper reports 98% accuracy for an arrhythmia model on a microcontroller. What do you ask?

Model answer

How the data were split. In one study accuracy was 98.07% overall but 80.73% when evaluated patient-independently at record level (Huang et al. 2026). Also: which classes, which database, and whether the number is for the quantized model running on the device.

Code lab

The lab builds each check on synthetic data and then audits real PhysioNet headers.

Part What you do Data
A Units in WFDB headers and an audit that catches header mistakes (Equation 1) synthetic ECGs
B Checksums, a manifest and a dataset version synthetic files
C Several labellers, majority votes and agreement (Equation 2) synthetic labels
D Re-identification risk and coarsening (Equation 3) synthetic registry
E Quantizing inputs and weights (Equation 4, Equation 5) synthetic beats
F A model card generated from the lab’s own results parts B and E
G Real headers, label-quality fields and an amplitude audit across databases PTB-XL and Challenge 2017 (download)
H Optional: dynamic int8 quantization in PyTorch synthetic beats
  • Python notebook: labs/python/m47_lab.ipynb. Run it from the course repository after pip install -r requirements.txt, or open it in Google Colab.
  • Data for part G: PTB-XL 1.0.3 (Wagner et al. 2020) (Creative Commons Attribution 4.0) and the PhysioNet/CinC Challenge 2017 training set (Clifford et al. 2017; Clifford et al. 2018) (Open Data Commons Attribution v1.0); only header files, the PTB-XL metadata table and a few recordings are downloaded, checked against the published SHA-256 lists, and the part is skipped without internet access.
  • Part H 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 stored value is 450 adu, the baseline 50 adu and the gain 200 adu/mV. What is the voltage?

Answer

\((450 - 50)/200 = 2\) mV (Equation 1).

S2. Three independent labellers are each right 90% of the time. How often is their majority right?

Answer

\(3(0.81) - 2(0.729) = 0.972\) (Equation 2).

S3. A table of 1,000 records falls into 250 groups, the smallest with 2 records. What are the largest and average risks?

Answer

Largest \(1/2 = 0.5\); average \(250/1000 = 0.25\) (Equation 3).

S4. An 8-bit unsigned quantizer covers −1 to 3. What are the scale and zero point?

Answer

\(S = 4/255 \approx 0.0157\) and \(Z = 0 - \operatorname{round}(-1/S) = 64\) (Equation 4).

S5. Going from 8 to 12 bits over the same range improves SQNR by about how much?

Answer

About \(4 \times 6 = 24\) dB (Equation 5).

S6. True or false: if a file opens without an error, its units are right.

Answer

False. A missing or wrong gain still gives a readable file, with every amplitude scaled (lab part A).

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.

  • PTB-XL’s structure. Wagner and colleagues describe how records, labels, metadata and the recommended folds fit together [Wagner et al. (2020); free full text]. This module: data formats and labels.
  • k-anonymity in practice. El Emam and Dankar compare the re-identification risk and information loss of k-anonymity and its variants on large health data sets [El Emam and Dankar (2008); free full text]. This module: re-identification risk.
  • De-identifying report text. Chambon and colleagues describe a pipeline that detects protected health information and replaces it with realistic surrogates [Chambon et al. (2023); free full text]. This module: de-identification.

Further reading

  • Formats and versions. PhysioNet (2026); Wilkinson et al. (2016).
  • Labels. Clifford et al. (2018), Wagner et al. (2020); agreement and label ceilings in m35.
  • Privacy. U.S. Department of Health and Human Services, Office for Civil Rights (2012), El Emam and Dankar (2008), Ghazarian et al. (2021), Chambon et al. (2023). Rules outside the United States differ; this module is not legal advice.
  • Deployment. Jacob et al. (2017), Wu et al. (2020), Huang et al. (2026), Faraone et al. (2021); federated training in m45.
  • Documentation. Mitchell et al. (2018), Gebru et al. (2018), Liang et al. (2024).
  • Back to the start of Part VII: m37, and the evidence ladder in m45.
  • AI competency areas. AI and society (privacy and data governance); machine learning (deployment). See the AI competency map.

References

Chambon, Pierre J, Christopher Wu, Jackson M Steinkamp, et al. 2023. “Automated Deidentification of Radiology Reports Combining Transformer and "Hide in Plain Sight" Rule-Based Methods.” Journal of the American Medical Informatics Association 30 (2): 318–28. https://doi.org/10.1093/jamia/ocac219.
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.
El Emam, Khaled, and Fida Kamal Dankar. 2008. “Protecting Privacy Using k-Anonymity.” Journal of the American Medical Informatics Association 15 (5): 627–37. https://doi.org/10.1197/jamia.M2716.
Faraone, Antonino, Halla Sigurthorsdottir, and Ricard Delgado-Gonzalo. 2021. “Atrial Fibrillation Detection on Low-Power Wearables Using Knowledge Distillation.” Annual International Conference of the IEEE Engineering in Medicine and Biology Society 2021: 6795–99. https://doi.org/10.1109/EMBC46164.2021.9630957.
Gebru, Timnit et al. 2018. Datasheets for Datasets. arXiv:1803.09010. https://doi.org/10.48550/arXiv.1803.09010.
Ghazarian, Arin, Jianwei Zheng, Hesham El-Askary, et al. 2021. “Increased Risks of Re-Identification for Patients Posed by Deep Learning-Based ECG Identification Algorithms.” Annual International Conference of the IEEE Engineering in Medicine and Biology Society 2021: 1969–75. https://doi.org/10.1109/EMBC46164.2021.9630880.
Hannun, Awni Y, Pranav Rajpurkar, Masoumeh Haghpanahi, et al. 2019. “Cardiologist-Level Arrhythmia Detection and Classification in Ambulatory Electrocardiograms Using a Deep Neural Network.” Nature Medicine 25 (1): 65–69. https://doi.org/10.1038/s41591-018-0268-3.
Huang, Jiayan, Chuansheng Wang, Antoni Grau, et al. 2026. “EdgeECG: A Lightweight Edge-Oriented Network with Dual Criterion Pruning for Real-Time ECG Arrhythmia Classification.” Physiological Measurement 47 (5). https://doi.org/10.1088/1361-6579/ae696a.
Jacob, Benoit et al. 2017. Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference. arXiv:1712.05877. https://doi.org/10.48550/arXiv.1712.05877.
Liang, Weixin, Nazneen Rajani, Xinyu Yang, et al. 2024. “Systematic Analysis of 32,111 AI Model Cards Characterizes Documentation Practice in AI.” Nature Machine Intelligence 6 (7): 744–53. https://doi.org/10.1038/s42256-024-00857-z.
Mitchell, Margaret et al. 2018. Model Cards for Model Reporting. arXiv:1810.03993. https://doi.org/10.48550/arXiv.1810.03993.
PhysioNet. 2026. Header: WFDB Header File Format (WFDB Applications Guide). https://physionet.org/physiotools/wag/header-5.htm.
U.S. Department of Health and Human Services. 2010. Workshop on the HIPAA Privacy Rule’s de-Identification Standard. https://www.hhs.gov/hipaa/for-professionals/privacy/special-topics/de-identification/2010-de-identification-workshop/index.html.
U.S. Department of Health and Human Services, Office for Civil Rights. 2012. Guidance Regarding Methods for de-Identification of Protected Health Information in Accordance with the Health Insurance Portability and Accountability Act (HIPAA) Privacy Rule. https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html.
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.
WFDB Python contributors. 2022. WFDB Python Package Documentation: Format Conversions (V4.0.0). https://wfdb.readthedocs.io/en/v4.0.0/convert.html.
Wilkinson, Mark D, Michel Dumontier, I Jsbrand Jan Aalbersberg, et al. 2016. “The FAIR Guiding Principles for Scientific Data Management and Stewardship.” Scientific Data 3: 160018. https://doi.org/10.1038/sdata.2016.18.
Wu, Hao et al. 2020. Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation. arXiv:2004.09602. https://doi.org/10.48550/arXiv.2004.09602.