Clinical records and laboratory results are among the most valuable training data that exist, and among the hardest to license well. This post sets out what can be licensed, how de-identification works, where it fails, what public deals say the data is worth, and where the rules stand as of October 2026.
What counts as usable data.
Not every record in a health system or a lab is a dataset. The ones that are share four properties: structure, depth over time, labels a model can learn against, and a provenance trail a buyer can defend.
| Data type | Example | Why models want it | Rights question |
|---|---|---|---|
| Structured EHR | Diagnoses, medications, labs, vitals, encounters over years | Longitudinal signal; outcomes linked to interventions | Is it PHI? Which de-identification path, and who signed it? |
| Clinical notes | Progress notes, discharge summaries, radiology reports | Natural language with expert reasoning; the scarcest text on earth | Free text carries identifiers that Safe Harbor lists never catch [2] |
| Imaging | CT, MRI, pathology slides with reports | Paired image and label at clinical quality | DICOM headers, burned-in text, and full-face images are identifiers [1] |
| Omics | Genomic, transcriptomic, proteomic results tied to phenotype | Genotype-to-phenotype pairs are the basis of most drug-target claims | Genomic data is hard to de-identify in principle; consent terms govern |
| Assay and screening data | Dose-response curves, plate reads, phenomic images, ADMET | Experimental ground truth that public databases lack [17] | Owned by the lab; sponsor and collaboration agreements may encumber it |
| Lab notebooks and protocols | Electronic notebook entries, SOPs, failed runs | Negative results and methods context, which are rarely published | Trade secret, joint inventions, and employee IP assignments |
| Operational data | Scheduling, claims, prior-authorization correspondence | Workflow models for administrative tasks | PHI in the payer context; business-associate chains |
De-identification, properly.
HIPAA offers two routes to data that is no longer protected health information, both at 45 CFR 164.514 [1].
Safe Harbor removes 18 categories of identifier: names; geographic units smaller than a state, with ZIP codes cut to three digits and zeroed where the area holds 20,000 people or fewer; all date elements other than year, with ages over 89 collapsed to “90 or older”; phone and fax numbers; email addresses; Social Security, medical record, health plan, account, and certificate numbers; vehicle and device identifiers; URLs and IP addresses; biometric identifiers; full-face photographs; and any other unique identifying number, characteristic, or code [1]. On top of the list, the covered entity must have no actual knowledge that what remains could identify someone [1][3].
Expert Determination keeps more of the data. A person with appropriate statistical or scientific knowledge documents that the risk is “very small that the information could be used, alone or in combination with other reasonably available information, by an anticipated recipient to identify an individual” [1]. OCR’s 2012 guidance says no specific degree or certification makes someone an expert, and it approves time-limited determinations, because the outside information an attacker could use changes over time [2][3]. Expiry does not make released data protected again; it means the next release needs a fresh look [3].
Two further tools sit beside them: a re-identification code, kept confidential and not derived from the patient’s information [1], and a limited data set, which strips direct identifiers but keeps dates and geography and can be shared for research under a data use agreement, though it remains PHI [1].
Where does it fail? Rocher, Hendrickx, and de Montjoye estimated in 2019 that 99.98% of Americans would be correctly re-identified in any dataset using 15 demographic attributes, and concluded that “even heavily sampled anonymized datasets are unlikely to satisfy the modern standards for anonymization set forth by GDPR” [4]. Sweeney’s group matched news stories to statewide hospital discharge data and uniquely re-identified 28.3% of Maine patients and 34% of Vermont patients in the raw release; applying Safe Harbor cut those figures to 3.2% and 10.6%, not to zero [5]. The counterweight is a 2011 systematic review by El Emam and colleagues, which found that most published attacks hit data that had not been de-identified to any standard, and that the one attack on standards-compliant health data re-identified 0.013% of records [6]. The honest reading: Safe Harbor is a floor, Expert Determination is a measurement, and the measurement must be repeated as public data grows.
Deals on the record.
Mayo Clinic and nference. STAT reported in 2020 that Mayo’s 16 data deals with technology companies had generated less than $5 million against more than $13 billion in annual revenue, and that patients had not been asked for deal-specific consent; the hospital said it was reconsidering its consent process [7]. Mayo invested $20 million in each of nference’s Series B and C rounds, and in March 2024 the two expanded a 12-year relationship built on Mayo Clinic Platform’s de-identified electronic health data [8].
Tempus, Pathos, and AstraZeneca. Tempus’s April 23, 2025 8-K states that Pathos will pay Tempus “data license fees of $200 million over a three-year period, including an upfront payment of $50 million,” in return for “a comprehensive de-identified multi-modal dataset for the sole purpose of assisting in the development and training” of an oncology foundation model; AstraZeneca pays a further $35 million, Tempus pays Pathos $35 million, and Pathos may pay up to half its fees in preferred stock [9]. This is the clearest public price on de-identified clinical data for model training we have found.
Truveta and Regeneron. On January 13, 2025 Regeneron announced a $119.5 million investment in Truveta’s Series C to launch a genome project sequencing up to ten million de-identified patient volunteers, drawing on a database of more than 120 million patients across 30 health systems [10].
Epic Cosmos. Epic’s research dataset reports 310 million patients, 2,358 hospitals, and 52.5 thousand clinics contributed by health systems on its software [11]. It is a commons, not a licensing market.
Flatiron. Roche bought the oncology EHR and real-world-evidence company for $1.9 billion in 2018, citing “regulatory-grade real-world evidence” as the reason [12].
NHS Federated Data Platform. Palantir’s consortium won a seven-year contract worth up to £330 million in November 2023 [13]. By May 2025 Palantir counted 72 of England’s 215 trusts as users [15]. By September 30, 2026, 38 Degrees had filed 44,000 Article 21 objections ahead of the contract’s early-2027 break clause [14]. A data structure that patients and clinicians do not trust is not a durable asset, whatever the headline value.
What a lab’s own data is worth.
AlphaFold was trained on roughly 170,000 publicly available protein structures from the Protein Data Bank [16]. That is sometimes read as proof that public data is enough. The opposite is closer to the truth: the models that work have a large, clean, experimentally grounded training set behind them, and most of biology has no such set. Basecamp Research’s 2025 preprint argues that public sequence databases “were established to aggregate results from academic experiments, not as training datasets for machine learning,” with 68% of sequence data in one major archive coming from five species [17].
That is why pharma pays for proprietary experimental data. Recursion’s December 2021 collaboration with Roche and Genentech carried $150 million upfront and up to 40 programs, each potentially worth more than $300 million in milestones, built on phenomic screening data that Recursion generates in its own automated labs [18][19]. Isomorphic Labs announced collaborations in January 2024 with Eli Lilly ($45 million upfront) and Novartis ($37.5 million upfront), together worth “nearly $3 billion” in potential milestones [20]. Those are drug-discovery deals, not data licenses, but what is being paid for is a model trained on data the partner cannot get elsewhere.
For a per-record view, L.E.K.’s 2023 survey of 200 healthcare-data buyers put the price of a patient record between roughly €50 for a simple episode and roughly €3,000 for a complex clinical dataset with detailed genomic information, with oncology and genomics records cross-linked to outcomes commanding more than €1,000 each [21]. Those are indicative figures from one survey, in euros, for clinical records rather than assay data; we know of no published per-record price for notebooks or screening data. What the deals above show is that value sits in paired, labeled, longitudinal data with a known origin, and a lab holding years of it has an asset it has probably never valued.
The rules in 2026.
HIPAA Security Rule. OCR published a proposed rule on January 6, 2025 that would remove the distinction between “addressable” and “required” safeguards and make encryption, multi-factor authentication, network segmentation, a technology asset inventory and network map, annual compliance audits, vulnerability scans every six months, and annual penetration tests mandatory [22][23]. The agency received close to 5,000 comments. As of October 2026 the rule is not final; the OMB Unified Agenda lists July 2027 for final action, pushed back from May 2026 [23]. The current Security Rule remains in force.
Enforcement. The inflation-adjusted penalty ceiling is $2,190,294 per violation category per year [24]. Risk-analysis failures remain the most common finding; 2026 settlements include OSF HealthCare ($552,250) and Assured Imaging ($375,000) [24].
Washington. The My Health My Data Act, signed April 27, 2023 and in force for most entities since March 31, 2024, requires a separate, signed authorization before consumer health data is sold, and makes violations enforceable under the state Consumer Protection Act, including by consumers [25][26]. It exempts HIPAA-protected information and de-identified data [25].
California. AB 254 extends the Confidentiality of Medical Information Act to reproductive and sexual health digital services, and AB 352 requires EHR systems to segregate and restrict access to abortion, gender-affirming-care, and contraception records, with the segregation provisions effective July 1, 2024 [27].
Consent and IRBs. Under the Common Rule, secondary research on existing data is exempt from consent when identities “cannot readily be ascertained,” the investigator does not contact or re-identify subjects, or when the use is already governed by HIPAA [28]. A properly de-identified dataset usually needs no new consent, but data collected under a specific protocol is bound by that protocol’s terms, and a sponsor’s study agreement often reaches further than the IRB does.
If you run a practice, a lab, or a biotech.
The deals above are for clinical and experimental data. Mikra takes patient and clinical records only through its healthcare track. On that track an independent de-identification provider de-identifies the records before Mikra receives anything, and the provider signs a Business Associate Agreement with the healthcare company. Mikra never receives protected health information. The track runs under a Healthcare and Legal Addendum to the agreement. The provider is not yet contracted, so the track is not open yet.
On the standard track, what we license from healthcare and biotech companies is their ordinary business records: mail, chat, documents, customer-relationship records, and finance, across the systems and years the company holds. A company that mainly provides medical or biotech services qualifies with 15 or more office-based employees in the United States or Canada. You keep ownership and keep using the records. We de-identify them before any customer receives anything. The risk is lower, not zero. You are paid a share of the revenue your data earns, and the exact terms are in your agreement, in writing, before any export. There is no upfront payment and no guaranteed amount: we pay only from what customers pay us.
If you hold clinical data and want to license it, the sections above describe what a buyer will ask. If you want to talk about either track, start a conversation.
Caveats.
Deal values are what the parties chose to disclose; upfront payments and milestone ceilings are different numbers, and most milestones are never paid. L.E.K.’s prices are 2023 survey indications, not transactions. Re-identification studies measure specific datasets against specific auxiliary data; they do not predict the risk of yours. The Security Rule proposal may change or never be finalized. We cited only two states.
This is not legal or medical advice.
This post summarizes public sources as of October 19, 2026. Whether a dataset is de-identified, whether a use needs consent or IRB review, and what an agreement permits are questions for qualified counsel and, where relevant, a de-identification expert working on your actual data.