A hospital data science team wants to model 30-day readmission risk after cardiac stent placement. They can license Optum's de-identified claims for roughly the cost of a senior hire, or they can spend nine months negotiating access to MIMIC and stitching together SEER records for free. Both paths are defensible. Choosing wrong burns a year and a budget, and sometimes a regulatory audit.
This lesson is about that choice: when to buy external data, when to build access to public sources, and how to judge fitness for regulated use.
Before comparing, know the players. These four appear in almost every biotech and medtech data conversation in the US.
SEER (Surveillance, Epidemiology, and End Results): a cancer registry run by the US National Cancer Institute. It covers roughly 48% of the US population across selected states and regions (as of recent NCI documentation, verify current coverage). Free, high quality, gold standard for cancer incidence and survival. Weak on treatment detail and comorbidities.
MIMIC (Medical Information Mart for Intensive Care): de-identified ICU data from Beth Israel Deaconess Medical Center in Boston, published via PhysioNet. Free, but single-institution. Extremely rich (vitals, labs, notes, waveforms) for tens of thousands of ICU stays. Great for method development, poor for national generalizability.
Optum claims: a large commercial and Medicare Advantage claims dataset sold by Optum (part of UnitedHealth Group). Covers hundreds of millions of lives over time. Claims tell you what was billed, not what clinically happened. Commercial license, priced per project or subscription.
"Build" here means investing staff time to access and engineer free or low-cost public data. "Buy" means licensing a commercial dataset that arrives cleaned and documented.
The decision is not mainly about money. It is about four axes.
Ask: does the source represent the population you will deploy on?
MIMIC is one Boston ICU. If you train a sepsis model there and deploy in a rural Texas hospital, expect drift. Optum skews toward commercially insured and Medicare Advantage members, so it underrepresents Medicaid and the uninsured. SEER covers cancer well but only in its registry regions.
A concrete rule: match the source population to your target population on age, payer mix, and geography before anything else.
Public data is free to download but expensive to use. Commercial data is expensive to license but cheap to use.
Worked example. Suppose licensing Optum for a one-year project costs an estimated 250,000 dollars (illustrative, not a quoted price). Building equivalent access from public sources needs 2 data engineers for 6 months.
Loaded cost per engineer/month (fully burdened) = 18,000 USD (estimate)
2 engineers x 6 months x 18,000 = 216,000 USD
Plus 40,000 USD compliance/legal review
Build total ≈ 256,000 USDThe headline costs are similar. The tiebreaker is what you get: the built version is bespoke and you own the pipelinepipelineAll active sales opportunities across the stages of the sales process, together with their combined potential value and probability of closing.Voir la définition complète →; the bought version is faster and vendor-supported but recurring. Model the second year, not just the first.
This is where regulated use lives or dies.
In the US, HIPAAHIPAAHealth Insurance Portability and Accountability Act, loi américaine imposant la protection des données de santé (PHI). Violations : amendes jusqu'à 1,9M$ par catégorie de violation. (Health Insurance Portability and Accountability Act) governs protected health information. De-identified data (under HIPAAHIPAAHealth Insurance Portability and Accountability Act, loi américaine imposant la protection des données de santé (PHI). Violations : amendes jusqu'à 1,9M$ par catégorie de violation.'s Safe Harbor or Expert Determination methods) can be used more freely. SEER, MIMIC, and Optum are all de-identified, but the standards and residual re-identification risk differ.
In Europe, GDPR (General Data Protection Regulation) is stricter. Health data is a "special category" requiring an explicit legal basis. "De-identified" under HIPAAHIPAAHealth Insurance Portability and Accountability Act, loi américaine imposant la protection des données de santé (PHI). Violations : amendes jusqu'à 1,9M$ par catégorie de violation. is not automatically "anonymized" under GDPR; if re-identification is reasonably possible, GDPR still applies. A US claims dataset licensed for a European trial can trigger a compliance gap here.
MIMIC requires a specific step: completing the CITI (Collaborative Institutional Training Initiative) human subjects research training and signing a data use agreement on PhysioNet before download. Free does not mean frictionless.
The PhysioNet credentialing process is a good concrete example of what a real DUA (data use agreement) looks like.
If your output supports an FDA submission, a label claim, or a reimbursement dossier, the bar rises sharply.
The FDA increasingly evaluates RWE (real-world evidence) drawn from RWD (real-world datareal-world dataRWD, données collectées en dehors des essais cliniques contrôlés : dossiers médicaux, claims d'assurance, données de dispositifs connectés, base des Real-World Evidence (RWE).) like claims and registries. Their guidance stresses data relevance and reliability: provenance, completeness, and whether the data captures the clinical concept you claim to measure.
MAUDE illustrates the trap. It is perfect for hypothesis generation ("are we seeing more reports of lead fractures?") and useless as a rate ("our failure rate is X"), because it has no denominator. Using MAUDE to claim a failure rate in a regulatory filing is a classic fitness error.
🎬 [VIDEO: "Real-World Evidence in Regulatory Decision-Making" - youtube.com - FDA and academic overview of how RWD sources are evaluated for submissions]
Score each candidate source 1 to 5 on the four axes, then weight by your use case.
| Axis | SEER | MIMIC | Optum | MAUDE |
|---|---|---|---|---|
| Coverage | 4 | 2 | 4 | 3 |
| Cost efficiency | 5 | 5 | 2 | 5 |
| Consent clarity | 5 | 4 | 3 | 4 |
| Regulated fitness | 4 | 2 | 4 | 2 |
These are illustrative judgments for a typical US analytics project, not universal truths. A method-development team would weight coverage low and richness high, flipping MIMIC's attractiveness. Build your own weights before you score.
For a regulatory submission, weight "regulated fitness" heavily and Optum or SEER pull ahead. For an early algorithm prototype, weight cost and richness and MIMIC wins.
Vérification des acquis
1. A team needs to model national cancer survival trends and requires the highest-quality incidence and survival data, even if treatment detail is thin. Which dataset is the best fit?
2. Why does the excerpt caution that claims data 'tell you what was billed, not what clinically happened'?
3. MAUDE is described as having 'no denominator.' What analytical limitation does this create?
4. Select ALL correct answers. A dataset is 'free' like MIMIC but described as poor for national generalizability. Which reasons support this concern?
Sélectionnez toutes les réponses correctes.
5. Select ALL correct answers. When deciding between licensing commercial data versus building access to public sources, which trade-offs does the lesson emphasize?
Sélectionnez toutes les réponses correctes.
Often the answer is neither pure buy nor pure build. It is linkage.
Tokenization lets you join datasets on de-identified patient tokenstokensA token is the basic unit of text that language models process, often a word fragment, whole word, or punctuation mark rather than a single character.Voir la définition complète → without exposing identities. Vendors like Datavant provide this brokering. You can link SEER cancer records to Medicare claims (the real, established SEER-Medicare linked dataset, maintained by NCI and CMS) to get incidence plus treatment plus cost in one view.
SEER-Medicare is worth knowing as a template: it solves SEER's weakness (thin treatment data) by joining to CMS (Centers for Medicare and Medicaid Services) claims. Coverage is limited to the Medicare-age population, so generalizability to younger patients is poor. Every linkage buys richness and costs generalizability somewhere.
Before signing any license or downloading any public set, get written answers to:
A one-page answer to these six questions prevents most expensive mistakes.