- A practice or hospital that is a HIPAA covered entity may license waveforms only once they are de-identified under 45 CFR 164.514, or with authorizations that state it is paid.
- ECG and EEG files carry identifiers inside them: EDF headers hold a patient code, sex, birthdate and name, and ECG exports carry acquisition times and device serial numbers.
- In a simulated attack, a model given one known ECG of a person found that person's ECG in anonymized datasets 79.2% of the time in a pool of 100 and 40.0% in a pool of 20,000.
- The Philips ECG AI Marketplace, launched in July 2025, distributes third-party ECG algorithms to hospitals; its launch release says nothing about buying practices' data.
Who may license ECG and EEG data, and under which rule?
It depends on who recorded the waveform and in what capacity. A resting ECG in a cardiology office, a long-term EEG in a hospital unit, a cloud-stored Holter study and a smartwatch tracing are governed by different rules. The table names each holder, the rule, its operative words and the plain meaning. The patient's consent, clinical or research, sits behind every row.
| Holder and rule or law | Operative words (quoted) | Plain meaning for a license | Source and date read |
|---|---|---|---|
| Practice, hospital or sleep and neurodiagnostic lab that is a HIPAA covered entity: 45 CFR 164.514 | Safe harbor removes, among others, "Device identifiers and serial numbers", "Biometric identifiers, including finger and voice prints" and "All elements of dates (except year)" | Waveforms may be licensed without patient authorization only once de-identified by safe harbor or expert determination | eCFR, read 20 November 2026 |
| The same holder: 45 CFR 164.502(a)(5)(ii) and 164.508(a)(4) | "a covered entity or business associate may not sell protected health information" except under an authorization that must "state that the disclosure will result in remuneration" | Identifiable waveforms may be licensed for payment only with each patient's authorization. Research disclosures under an IRB or privacy board waiver or a limited data set agreement, paid only a "reasonable cost-based fee" to prepare and send the data, are not a sale; a fee above cost is | eCFR, read 20 November 2026 |
| Device maker or ECG management vendor holding hospital data: business associate under 45 CFR 160.103 | A business associate may de-identify PHI "only to the extent such activity is authorized by their business associate agreement" | The vendor may license customer waveforms only if its BAA and contract allow de-identification and its own use of the result | eCFR and HHS, read 20 November 2026 |
| Device makers generally: FDA | Data-transfer software that only stores or displays device data is "not subject to FDA regulatory requirements applicable to devices" | FDA regulates device safety and effectiveness. We found no FDA rule that decides who owns device data or who controls its licensing | fda.gov, read 20 November 2026 |
| Consumer wearable or app maker outside HIPAA: Washington My Health My Data Act, RCW 19.373 | Consumer health data includes "Bodily functions, vital signs, symptoms, or measurements"; selling it requires "valid authorization from the consumer" | Selling consumer ECG data about Washington consumers needs a signed authorization, separate from the consent to collect it | Washington State Legislature, read 20 November 2026 |
| Consumer app maker outside HIPAA: FTC Health Breach Notification Rule, 16 CFR 318 | A breach "occurs as a result of a data breach or an unauthorized disclosure" | For an app that is a personal health record vendor, sharing users' data without their authorization can trigger the rule's notice duties. It does not apply to HIPAA covered entities, or to a business associate when acting as one | eCFR, read 20 November 2026 |
| University or research lab: Common Rule, 45 CFR 46 | Information is identifiable when identity "is or may readily be ascertained by the investigator or associated with the information" | Research recordings reused for new research need consent, broad consent, an exemption or an IRB waiver, and the original consent wording governs | eCFR, read 20 November 2026 |
| Any US holder with a foreign licensee: 28 CFR 202 | Bulk personal health data is data on "more than 10,000 U.S. persons"; the definition includes "vital signs" | Data on more than 10,000 US persons sent to the same foreign person in one deal, or across deals in 12 months, is bulk even when de-identified. A license to the six countries of concern or covered persons is prohibited under 28 CFR 202.301; other foreign licensees of bulk data need an onward-transfer clause, and known or suspected breaches must be reported | eCFR, read 20 November 2026 |
| Holder of data on people in the EU: GDPR, Articles 6, 9 and 44 | Processing of "data concerning health" and of biometric data used for unique identification is prohibited unless an Article 9(2) condition applies; member states "may maintain or introduce further conditions" | EU patients' waveforms need an Article 6 basis and an Article 9(2) condition, such as explicit consent for specified purposes or research under Article 89(1), plus any member state conditions under Article 9(4), and a Chapter V transfer mechanism for licensees outside the EEA. Data de-identified to HIPAA's standard is not anonymous for that reason alone (Recital 26) | EUR-Lex, read 20 November 2026 |
| Consumer EEG or neurotechnology maker outside HIPAA: state neural data laws in California, Colorado, Connecticut and Montana | California: "Neural data" means "information that is generated by measuring the activity of a consumer's central or peripheral nervous system, and that is not inferred from nonneural information" | EEG held outside HIPAA can be sensitive data under these laws, each with its own scope, thresholds and dates, set out in the section on neural data below | Each state's official site, read 20 November 2026 |
Sources: 45 CFR 164.514, 164.502, 164.508, HHS de-identification guidance, FDA on data systems, RCW 19.373, 16 CFR 318.2, 45 CFR 46.102, 28 CFR part 202, GDPR and California Civil Code 1798.140, read on 20 November 2026. This page is general information, not legal advice.
What is waveform data, and why do AI companies want it?
Waveform data is the raw electrical signal a device records over time, sampled many times a second, as opposed to the one-line interpretation in a chart. A resting 12-lead ECG is usually a 10-second recording on 12 channels; the MIMIC-IV-ECG dataset describes its tracings as 12-lead, 10 seconds long and "sampled at 500 Hz". Holter and patch monitors record for days. EEG records many scalp channels for minutes to days, often with video and technician annotations.
AI developers train models to read what a clinician cannot see in the trace. Philips described one such model in July 2025: an algorithm that "analyzes a standard 12-lead resting ECG to identify reduced ejection fraction, an early indicator of heart failure". Models like that need tracings linked to later diagnoses, imaging and outcomes, which is why a waveform with its follow-up is more useful than a waveform alone.
Large public sets set the baseline any license is compared against. PTB-XL on PhysioNet holds "21799 clinical 12-lead ECG records of 10 seconds length from 18869 patients" under a Creative Commons Attribution license. MIMIC-IV-ECG offers "approximately 800,000 diagnostic electrocardiograms across nearly 160,000 unique patients" under the Open Database License. The CHB-MIT database is "a collection of EEG recordings of 22 pediatric subjects with intractable seizures". A licensed archive has to offer something these do not: scale in a specific population, linked outcomes, rare conditions or recent devices.
Are ECG and EEG waveforms protected health information?
Yes, when a covered entity or business associate holds them and they can be tied to a patient, which is almost always true of files exported from clinical systems. The harder question is whether a waveform stripped of every header field is still identifiable.
Published research says it can be. A 2025 study in European Heart Journal Digital Health trained a model on a hospital dataset of 864,283 ECGs and concluded that "Anonymized ECGs retain information that may facilitate subject re-identification". Its test was a simulation: "In a simulated ‘motivated intruder’ test", the model was given one reference ECG of a known person and searched pools of 100 to 20,000 ECGs in which only one other ECG was that person's, and "it achieves success rates of 79.2%, 62.6%, 45.0%, and 40.0%, respectively". The study did not re-identify real people from a released dataset. A 2023 systematic review in The Lancet Digital Health of re-identification from wearable signals, with ECG and EEG among the most studied, found that "Correct identification rates were typically 86–100%".
HIPAA's safe harbor list names "Biometric identifiers, including finger and voice prints" and "Any other unique identifying number, characteristic, or code". Whether a raw ECG or EEG trace falls under either is not settled by the text, and we found no HHS statement on it. Safe harbor also fails if the holder has "actual knowledge" the remaining data could identify someone. For raw waveforms, an expert determination that weighs the published re-identification results is the safer route. See the 18 HIPAA identifiers guide for the full list.
Which state laws treat EEG as neural data?
EEG measures the brain's electrical activity, so state laws on neural data can reach EEG held outside HIPAA, for example by a consumer headset or wellness app maker. We read these four on the states' official sites. Other states may have similar laws, and this list may be incomplete.
- California: the consumer privacy act lists "A consumer's neural data" as sensitive personal information, added by SB 1223 (approved 28 September 2024) and so in force from 1 January 2025. It applies to businesses that meet a threshold, such as annual revenue over $25 million (as adjusted) or buying, selling or sharing data on 100,000 or more consumers or households. Consumers may direct a business to limit its use of sensitive personal information. PHI held by HIPAA covered entities and business associates is outside the act.
- Colorado: since 7 August 2024 the privacy act has counted "Biological data" as sensitive data, and biological data "includes neural data" when it is used or intended to be used "for identification purposes". A controller "shall not process or sell a consumer's sensitive data without first obtaining the consumer's consent". The act reaches controllers that process data on 100,000 Colorado consumers a year, or 25,000 with revenue from selling data, and excludes HIPAA PHI.
- Connecticut: from 1 July 2026, Public Act 25-113 made "Neural data", meaning "any information that is generated by measuring the activity of an individual's central nervous system", sensitive data. A controller may not process it without the consumer's consent or sell it without consent. The act reaches persons doing business in or targeting the state that processed data on 35,000 consumers in the prior year, and persons that process any sensitive data or offer personal data for sale.
- Montana: since 1 October 2025 its Genetic Information Privacy Act has covered "neurotechnology data" and requires "separate express consent" for each transfer to a third party, naming it, and express consent for any sale. HIPAA PHI is excluded only where separate informed consent was obtained and the entity follows the act's policies.
How are EEG and ECG files de-identified before licensing?
Start with the file formats, because identifiers sit in fixed places. EEG is often stored in EDF or EDF+. The EDF header has an 80-character "local patient identification" field, and the EDF+ specification fills it with "the code by which the patient is known in the hospital administration", sex, birthdate and the patient's name. Its recording field carries the start date and "The hospital administration code of the investigation, i.e. EEG number or PSG number", plus codes for the technician and equipment. ECG exports come as vendor XML, DICOM or PDF files. DICOM's de-identification profile lists "Device Serial Number" and "Acquisition DateTime" among the attributes to handle, and vendor XML and PDF exports need the same field-by-field check.
- Blank or replace every patient field in the header. A re-identification code must not be derived from the patient, used for any other purpose or disclosed with its mechanism (45 CFR 164.514(c)).
- Remove identifiers of relatives, employers and household members, geographic units smaller than a state (such as site ZIP codes in headers), and ages over 89, which safe harbor groups into one category of 90 or older.
- Remove device serial numbers and site or cart identifiers, which safe harbor lists as "Device identifiers and serial numbers".
- Reduce dates to the year, or shift all of a patient's dates by one offset if an expert determination allows it, so intervals between recordings survive.
- Strip free text: technician notes, EEG event annotations and ECG interpretation statements can carry names, rooms and dates.
- Remove synchronized video and audio from EEG studies unless the license and the de-identification method cover faces and voices.
- Keep what makes the signal useful: sample rate, lead or electrode montage, filter settings, units and calibration, which identify nobody.
Imaging studies need a separate process; the medical imaging licensing guide and the PACS guide cover DICOM images.
Who owns ECG data: the practice, the hospital or the device maker?
For tracings recorded in clinical care, the practice or hospital that keeps the medical record is usually the one entitled to decide on a license, under state medical records law and its own policies. The who owns patient records guide sets out the state rules. A practice that reads studies for a hospital under contract may hold copies it is not entitled to license; the reading agreement decides.
Device makers and ECG or EEG management vendors usually hold hospital data as business associates. HHS says a covered entity may use a business associate to de-identify data "only to the extent such activity is authorized by their business associate agreement". A vendor that wants to license data it stores for customers needs that permission, and often a separate license, from each customer. FDA rules do not change this: we found no FDA provision that assigns ownership of device data.
Consumer devices are different. A smartwatch maker that sells directly to consumers is usually outside HIPAA, so its terms of use, privacy notice and state laws such as Washington's My Health My Data Act decide. Who owns the data in that case is a question of contract and consumer law, and the answer differs by company and state.
Can a cardiology or neurology practice license ECG data for AI training?
Yes, if it owns the records, de-identifies them to HIPAA's standard and has no contract that stops it. We help you license your data. Origin Data Partners is an independent data partner that helps business owners value and license their company records: the company pays Origin nothing, and a buyer may pay Origin when a deal completes. In practice the work is mostly inventory and contracts.
- Count what you hold: resting ECGs, Holter and patch studies, stress tests, routine and long-term EEG, by year and device, and whether raw signals or only PDFs survive.
- Check where the files live and who holds them: your own server, a hospital system or a vendor cloud, and what each contract says about secondary use.
- Identify linked data you can add after de-identification, such as echo results, diagnoses and outcomes, which raise the value of a tracing.
- Choose the de-identification route, usually an expert determination for raw waveforms, and document it (see HIPAA de-identification).
- Separate any data from research studies, which follows its consent forms and IRB terms rather than your clinical rules.
When you have the counts, check what your waveform archive may be worth. For the documents buyers request, see health data licensing requirements.
What is the Philips ECG AI marketplace?
It is a channel for selling ECG algorithms to hospitals, not a program for buying ECG data. Philips announced the Philips ECG AI Marketplace on 17 July 2025 as "a platform that gives cardiac care teams access to multiple vendor offerings all in one central location". It said the marketplace "enables U.S. hospitals and health systems to integrate validated algorithms into their workflow and existing Philips infrastructure", and that Anumana's low ejection fraction algorithm was the first third-party tool on it.
The release says nothing about Philips or the marketplace buying data from practices, and the Philips pages we read do not offer to. Algorithm developers on such a platform do need training data, which is where waveform holders come in, but any license would be with the developer under its own terms. No buyer has published a price for ECG or EEG waveform data.
How we checked this
We read every source on this page on 20 November 2026. The rules come from their official texts: the eCFR (45 CFR parts 46, 160 and 164; 16 CFR part 318; 28 CFR part 202), HHS de-identification guidance, FDA's page on device data systems, the Washington, California, Colorado, Connecticut and Montana legislatures' sites and EUR-Lex for the GDPR. File-format facts come from the EDF and EDF+ specifications and DICOM PS3.15. Dataset facts come from PhysioNet project pages. Re-identification findings come from peer-reviewed studies in European Heart Journal Digital Health and The Lancet Digital Health. Market facts come from Philips's own release.
What we could not settle: whether a raw ECG or EEG trace is a biometric identifier under HIPAA safe harbor, which HHS has not addressed; the secondary-use terms in specific device makers' and vendors' contracts, which are not public; and any price for waveform data, because no buyer has published one. This page does not cover consumer health data or neural data laws in states not named here, FTC Act enforcement beyond the breach rule, 42 CFR part 2 for substance use treatment records, children's data and pediatric research rules, or the European Health Data Space Regulation.
Origin Data Partners works with companies that own data and helps get it bought. We are independent of the services on this page, and a buyer may pay us when a deal closes (how Origin is paid). See what your data is worth.
Common questions
Can EEG recordings be licensed to AI companies?
Yes, by the hospital, practice or lab that holds them, once they are de-identified under HIPAA or covered by patient authorizations, and when no vendor contract stops it. EEG adds steps: EDF headers carry the patient's name and birthdate, long-term studies often include video, and technician annotations hold free text. Research EEG follows its consent forms and IRB terms instead of clinical rules.
Are smartwatch ECGs covered by HIPAA?
Usually not while the maker holds them for a consumer, because a device sold straight to consumers is not a HIPAA covered entity. They can become PHI when a patient sends them to a provider who keeps them in the record. Outside HIPAA, the maker's terms and state laws such as Washington's My Health My Data Act apply. The FTC's Health Breach Notification Rule reaches personal health record vendors and related entities and sets breach notice duties.
Can a device maker license ECGs its devices recorded at hospitals?
Only if its contracts with those hospitals allow it. A maker that stores or processes the data for hospitals is usually a business associate, and HHS says it may de-identify data only as the business associate agreement authorizes. The agreement sets what else, if anything, the maker may do with the data. Owning the device does not give the maker the patients' tracings.
Is a free public ECG dataset a substitute for licensed data?
Partly. PTB-XL and MIMIC-IV-ECG on PhysioNet give anyone hundreds of thousands of tracings under open licenses, and credentialed sets add terms such as a promise not to identify anyone. A licensed archive has to offer what those sets lack: specific populations, rare conditions, current devices, long recordings and tracings linked to later outcomes.
Does the DOJ bulk data rule apply to ECG data?
It can. Under 28 CFR part 202, personal health data includes vital signs and bodily functions. Data on more than 10,000 US persons supplied to the same foreign person in one deal, or across deals in 12 months, is bulk even when de-identified. Licensing it to six countries of concern or covered persons is prohibited (202.301), and a vendor in one is a restricted transaction (202.401). A domestic license is not barred unless the licensee is a designated covered person.
Do Holter and event monitor recordings count as protected health information?
Yes, when a covered entity or its business associate holds them in identifiable form. Monitoring services that bill insurers and keep reports for ordering physicians are typically in that position. The recordings carry patient demographics, device serial numbers and dates, all of which safe harbor requires removing, and long recordings give an expert more signal to weigh for re-identification risk.
Can we keep sample rate and lead information after de-identification?
Yes. Technical metadata such as sample rate, lead or electrode montage, filter settings, amplitude units and calibration describes the recording, not the patient, and is what makes a waveform usable for training. Keep it. Remove the device serial number, dates other than the year (or shift them under an expert determination), and every free-text field that might name a person, room or date.
Sources and further reading
- eCFR: 45 CFR 164.514, de-identification of protected health information
- eCFR: 45 CFR 164.502, including (a)(5)(ii), sale of protected health information
- eCFR: 45 CFR 164.508, uses and disclosures for which an authorization is required
- eCFR: 45 CFR 160.103, definitions (covered entity, business associate, health information)
- HHS Office for Civil Rights: guidance on HIPAA de-identification methods
- FDA: Medical Device Data Systems
- Washington State Legislature: My Health My Data Act, RCW 19.373
- California Legislative Information: Civil Code 1798.140, CCPA definitions (sensitive personal information, neural data)
- California Legislative Information: Civil Code 1798.121, right to limit use of sensitive personal information
- California Legislative Information: SB 1223 (2024), neural data, chaptered text
- California Legislative Information: California Constitution, article IV, section 8 (when statutes take effect)
- Colorado Office of Legislative Legal Services: Colorado Revised Statutes 2026, title 6 (Colorado Privacy Act, 6-1-1301 to 6-1-1314)
- Connecticut General Assembly: Public Act 25-113 (SB 1295), amending the Connecticut Data Privacy Act (PDF)
- Montana Code Annotated 2025: 30-23-104, consumer genetic or neurotechnology data, consent
- Montana Code Annotated 2025: 30-23-102, definitions (neurotechnology data)
- Montana Code Annotated 2025: 30-23-103, exceptions
- Montana Legislature: effective dates by chapter number, 2025 (PDF)
- eCFR: 16 CFR 318.2, Health Breach Notification Rule definitions
- eCFR: 16 CFR 318.1, Health Breach Notification Rule scope
- eCFR: 45 CFR 46.102, definitions (identifiable private information)
- eCFR: 28 CFR part 202, access to US sensitive personal data by countries of concern
- EUR-Lex: Regulation (EU) 2016/679 (GDPR)
- European Heart Journal Digital Health (2025): Siamese neural network ECG can re-identify anonymized data (PMC)
- The Lancet Digital Health (2023): does deidentification of wearable data give a false sense of security? A systematic review
- EDF specification: header record
- EDF+ specification: patient and recording identification fields
- DICOM PS3.15 Annex E: attribute confidentiality profiles
- PhysioNet: PTB-XL, a large publicly available ECG dataset (version 1.0.3)
- PhysioNet: MIMIC-IV-ECG diagnostic ECG matched subset (version 1.0)
- PhysioNet Credentialed Health Data License 1.5.0
- PhysioNet: CHB-MIT Scalp EEG Database (version 1.0.0)
- Philips: Philips launches ECG AI Marketplace (17 July 2025)
