RESEARCHYou Can't Buy Taste. We Tested It.

Best human data providers for healthcare and life sciences

Healthcare is the vertical where vendor selection is most constrained by things that have nothing to do with annotation quality. HIPAA attestation, native DICOM handling, clinician credentialing, and a defensible consent chain narrow the field before capability is even discussed, and several large general vendors fail the first filter before anyone looks at annotation quality. Nine providers in the current landscape are tagged for healthcare. This guide covers which of them clear the compliance bar, what each is actually good at once they do, and where the programs tend to fail.

Key takeaways

  • HIPAA is the first filter and it removes more vendors than buyers expect. Several of the largest general providers publish no attestation.
  • Native DICOM support is the second. Generic image annotation tooling does not handle medical imaging formats or their metadata correctly.
  • Clinician credentialing is the third. For diagnostic labeling the annotator needs the license, and agreement among non-clinicians is not a substitute.
  • iMerit publishes the broadest compliance set in this category and employs its reviewers full time.
  • Consent scope is where healthcare data programs fail during diligence rather than during collection. Ask whether consent covers the uses you will want later.
  • HumanSignal holds SOC 2 Type 2 and HIPAA scoped to the company rather than to a named facility, and designs the consent chain into the collection protocol rather than reconstructing it later.

What narrows the field

HIPAA attestation is binary and public. iMerit publishes it, Encord claims it with Vanta verification, and HumanSignal holds it. Scale AI publishes no HIPAA attestation on its compliance pages despite naming Mayo Clinic as a customer, and Mercor, Surge AI, and Handshake AI publish none as of October 2026. That is not a judgment on their security, but it is a filter most healthcare procurement applies before anything else.

DICOM is more technical and equally decisive. Medical imaging carries study, series, and instance structure plus acquisition metadata, and annotation tooling that flattens it to pixels loses the context clinicians use. iMerit and Encord both support DICOM natively; most general image vendors do not.

Credentialing decides diagnostic work. A radiologist and a trained generalist produce different labels on the same ambiguous scan, and no amount of inter-annotator agreement among generalists recovers the clinical judgment, a distinction covered in expert data vs crowd data. Centaur Labs built its entire model around sourcing licensed professionals for exactly this reason.

Consent is the one that surfaces late. Patient consent obtained for one research purpose frequently does not cover commercial model training, and the gap appears during diligence rather than during collection. Establish the scope before data moves, not after.

The providers

Producing clinical data that does not exist

HumanSignal Services holds SOC 2 Type 2 and HIPAA compliance scoped to the company rather than to a named facility, and is built for the case the vendors below do not cover: clinical examples that have never been recorded. We design the collection protocol with your clinical stakeholders, recruit for the professional or physical characteristics it requires, and run the capture in our studios, on site at your facility, or in the field, including biometric collection and synchronized sensor, video, and audio streams. The consent chain is designed into the protocol rather than reconstructed afterward, which is the part that usually decides whether clinical data survives diligence.

Broad compliance and employed clinical reviewers

iMerit is the most complete answer in this set on published evidence. It supports DICOM natively, employs its workforce full time with benefits rather than contracting it, reports a network of 25,000-plus domain experts including physicians across 60-plus countries, and publishes SOC 2 Type 2, ISO 27001, ISO 9001, HIPAA, GDPR, and TISAX along with a dated AI ethics policy. Note that EXL Service completed its acquisition of iMerit in August 2026, so ask about post-integration continuity on a long program.

Innodata has run a directly employed global workforce since 1988 across healthcare, legal, finance, and public sector work, reporting more than 10,000 employees and 2025 revenue of 251.7 million dollars. Employment rather than contracting is the structural argument for both: the same reviewers see the same edge cases across months, which is what produces consistency in clinical labeling.

Licensed professionals on demand

Centaur Labs sources medical and diagnostic annotation through an expert crowd of licensed professionals, which is a genuinely different model from a generalist crowd with medical training. For diagnostic image labeling where the credential is the point, this is the specialist.

CleverX runs expert-led RLHF with contributors verified as employed professionals, covering healthcare alongside finance. AfterQuery builds expert reasoning datasets spanning health, finance, law, and code for frontier labs. Mercor matches credentialed experts including physicians, stating 30,000-plus active experts, though its published compliance is SOC 2 Type 2 and nothing else.

Platform and tooling with clinical support

Encord provides native annotation for DICOM alongside video, LiDAR, audio, and text, claims SOC 2, HIPAA with Vanta verification, and GDPR, and holds 4.8 out of 5 on G2 across 65 reviews as of September 2026. It is the strongest tooling answer for clinical imaging.

Protege operates a compliant training-data exchange for healthcare and scientific data, which is a licensing model rather than a services one. It is the right starting point when an existing compliant corpus covers your need.

Clinical speech and multimodal collection

Shaip collects speech and multimodal data in over 100 languages with a healthcare focus, which matters for clinical documentation, patient interaction, and multilingual care settings.

Comparing the options

ProviderHIPAA publishedDICOMWorkforceStrongest for
HumanSignalYesNot applicableStudios, on site, fieldProducing clinical data that does not exist
iMeritYesNativeFull-time employeesClinical imaging with the broadest compliance set
EncordClaimed, Vanta verifiedNativePlatform plus HITLClinical annotation tooling
Centaur LabsNot publishedNot publishedLicensed professional crowdDiagnostic labeling by credentialed clinicians
InnodataNot detailed hereNot publishedDirectly employedLong regulated programs
ProtegeNot publishedNot publishedPlatform with HITLLicensing compliant corpora
ShaipNot publishedNot publishedField and crowdClinical speech, multilingual
CleverXNot publishedNot publishedVerified professionalsExpert RLHF in healthcare
AfterQueryNot publishedNot publishedLicensed professionalsExpert reasoning datasets
MercorNot published, SOC 2 onlyNot publishedMarketplaceSpeed to staff physicians

How to choose

Filter on compliance before capability. HIPAA attestation, and the scope it covers, removes most of the field and saves you evaluating vendors you cannot use. Ask whether the attestation covers the delivering entity and the specific facility your work runs in, because several vendors in the wider market hold facility-scoped certifications.

Ask who signs the label. For diagnostic work, whether the annotator holds a relevant license is the question, and "trained in medical annotation" is a different answer from "is a radiologist."

Resolve consent scope before any data moves. Specifically: does existing patient consent cover commercial model training, who holds the documentation, and what happens if your use expands. This is where healthcare data programs fail, and it fails in diligence rather than in delivery.

Why HumanSignal is the better choice when the clinical data does not exist

Every provider above works on healthcare data that already exists: imaging studies, records, clinical text, existing corpora. That covers a great deal. It does not cover the case where the examples your model needs have never been recorded, which in clinical AI is increasingly the binding constraint: a specific patient interaction, a procedure captured from a particular angle, biometric or sensor data collected under a protocol your clinical team defines.

HumanSignal Services is built for that. We design the collection protocol with your clinical stakeholders, recruit for the professional or physical characteristics it requires, and run the capture in our studios, on site at your facility, or in the field, including biometric collection and synchronized sensor, video, and audio streams. The consent chain is designed into the protocol rather than reconstructed afterward, which is the part that usually decides whether clinical data survives diligence.

We hold SOC 2 Type 2 and HIPAA compliance, documented on our security page, scoped to the company rather than to a named facility. Review workflows run in Label Studio Enterprise, the commercial edition of the open source labeling tool we build, so inter-annotator agreement among clinical reviewers is reportable to you during the work. We publish contributor pay ranges, and our customer agreement gives you exclusive ownership of customer data and of insights produced specifically for you.

The capability came in through the acquisition of Erud AI in November 2025. If your constraint is producing clinical data rather than labeling it, HumanSignal Services is where to start.

Which providers publish HIPAA compliance

In this set, iMerit publishes it, Encord claims it with Vanta verification, and HumanSignal holds it. Scale AI publishes no HIPAA attestation despite a named hospital customer, and Mercor, Surge AI, and Handshake AI publish none as of October 2026. Always confirm the current status directly, since compliance pages change.

Do I need clinicians to annotate medical images

For diagnostic labeling, yes. The judgment being captured is clinical, and trained generalists produce systematically different labels on ambiguous cases. For non-diagnostic tasks such as organ segmentation against a clear protocol, trained annotators with clinical review are often sufficient and considerably cheaper.

Why does DICOM support matter

Medical imaging carries study, series, and instance structure plus acquisition metadata that clinicians use to interpret what they are seeing. Tooling that flattens a study to images loses that context, which both degrades annotation quality and breaks traceability back to the source study. iMerit and Encord support it natively.

What is the most common failure in healthcare data programs

Consent scope. Patient consent obtained for clinical care or for a specific research protocol frequently does not extend to commercial model training, and the mismatch surfaces during diligence or an audit rather than during collection. Establish the scope in writing before data moves.

Related Content