Establishing Lawful Basis First
Before any record is touched, the question is whether this data may be used for model development at all. Consent language, contract terms and applicable regulation decide it, and the answer is sometimes no.
Custom Software
Clinical Data Annotation and Healthcare AI Training Data
Clinical data annotation is the process of labelling healthcare records so they can train or evaluate a machine learning model. The work covers de-identification under HIPAA Safe Harbor or Expert Determination, annotation schema design, clinician-led labelling, inter-annotator agreement measurement and quality assurance. It is performed on datasets the client already holds and controls, under a Business Associate Agreement, with provenance documented for every record.
The hard part is not the labelling. It is establishing that the data may lawfully be used for this purpose, de-identifying it in a way that survives scrutiny, and producing labels consistent enough that a model trained on them generalizes. Taction has built HIPAA-compliant healthcare software since 2013. We annotate data our clients own. We do not source, broker or resell clinical datasets.
Tell Us Your Requirements
Our experts are ready to understand your business goals.
Trusted Partners
Trusted by Industry Leaders Worldwide
Recognition
Awards & Recognitions
What Clinical Data Annotation Actually Involves
Teams arrive with a model architecture and a dataset and assume annotation is a throughput problem. It is a definitions problem first. Two clinicians labelling the same discharge summary will disagree on whether a mention constitutes a diagnosis, an historical condition or a ruled-out possibility, and no amount of annotator volume resolves that. The schema, the guideline document and the adjudication process determine whether the resulting labels are usable. Everything else is execution against those decisions.
De-identification Before Annotation
Annotation should operate on de-identified data wherever the task permits. Where it does not, the work runs under a Business Associate Agreement with identified data in a controlled environment.
Schema and Guideline Design
Defining what a label means, what the edge cases are, and how ambiguity is resolved. This document is the single largest determinant of annotation quality.
Clinician-Led Labelling
Clinical judgment cannot be crowdsourced. Annotation requiring diagnostic interpretation needs annotators with the relevant clinical background, which changes both cost and throughput.
Agreement Measurement and Adjudication
Multiple annotators on overlapping samples, agreement measured statistically, disagreements adjudicated and the guideline updated. Without this, quality is asserted rather than known.
Provenance and Auditability
Which record, from which source, annotated by whom, under which guideline version. Required for regulatory scrutiny and for debugging a model that behaves unexpectedly.
De-identification: Safe Harbor Versus Expert Determination
HIPAA provides exactly two routes to de-identified status, and the choice between them shapes both the cost of the work and the utility of the resulting dataset. Safe Harbor is prescriptive and cheap to execute but strips information models frequently need, particularly dates and geography. Expert Determination is more expensive and slower but can retain analytically valuable fields where a qualified expert can document that re-identification risk is very small. Choosing Safe Harbor by default is common and frequently the wrong call for a model that depends on temporal sequence.
01
How Safe Harbor Works
Remove all eighteen specified identifier categories and have no actual knowledge that residual information could identify an individual. Mechanical, defensible and requires no expert sign-off.
02
What Safe Harbor Costs You Analytically
All dates other than year are removed, which destroys temporal sequence. Geography collapses to three-digit ZIP at best. For many clinical models this is the difference between viable and useless.
03
How Expert Determination Works
A person with appropriate statistical and scientific expertise determines that re-identification risk is very small, and documents the methods and results supporting that determination.
04
When Expert Determination Is Worth It
Where the model needs dates, intervals, granular geography or rare conditions retained. The determination cost is usually recovered in model performance and dataset longevity.
05
Limited Data Sets and Data Use Agreements
A third path retaining dates and some geography, permitted for research, public health and operations under a data use agreement rather than full de-identification.
06
Re-identification Risk Does Not Stay Fixed
A determination made against today’s auxiliary data can weaken as external datasets grow. Periodic reassessment belongs in the governance plan.
The Eighteen Safe Harbor Identifiers
Safe Harbor is mechanical, which makes it auditable, and it is also where most de-identification failures occur, because free text does not respect field boundaries. A clinical note that mentions a patient’s employer, their rare occupation or the name of the nursing home they transferred from carries identifying information that no structured-field scrubber will catch. Automated de-identification gets most of the way and the residual requires human review, which is a cost line teams routinely omit from their plan.
Direct Identifiers
Names, Social Security numbers, medical record numbers, health plan beneficiary numbers, account numbers, and certificate or licence numbers. The easiest category and the one automation handles reliably.
Contact Information
Telephone and fax numbers, email addresses, URLs and IP addresses. Straightforward in structured fields and frequently missed inside narrative text and scanned documents.
Geographic Subdivisions
Everything smaller than a state, with a conditional exception for the first three ZIP digits where the population of that area exceeds twenty thousand.
Dates
All date elements other than year, including admission, discharge and death dates, plus aggregation of all ages over eighty-nine into a single category.
Device, Vehicle and Biometric Identifiers
Device serial numbers, vehicle identifiers and plate numbers, and biometric identifiers including finger and voice prints. Common in device-derived and imaging datasets.
Images and Any Other Unique Code
Full-face photographs and comparable images, plus any other unique identifying number, characteristic or code. The final catch-all is the clause that requires human judgment rather than a rule.
Annotation Types for Healthcare AI
Annotation work divides into recognizable task types, and cost per record varies across them by more than an order of magnitude. A binary document classification is fast. Span-level clinical entity annotation with attribute assignment by a qualified clinician is not. Scoping an annotation project means pricing each task type separately rather than quoting a blended per-record rate, because the mix usually shifts once the guideline is written and the genuinely ambiguous cases become visible.
Clinical Named Entity Recognition
Marking conditions, medications, procedures, anatomy and devices in narrative text, typically with attributes for negation, uncertainty, temporality and experiencer.
Document and Section Classification
Assigning document types, section boundaries and note categories. Faster and cheaper, and often the prerequisite for everything downstream.
Terminology Normalization
Mapping annotated mentions to SNOMED CT, LOINC, RxNorm or ICD-10 codes. Slower than marking spans and considerably more valuable for interoperability.
Relation and Event Annotation
Linking entities to each other, such as a medication to its dosage and indication, or a symptom to its onset. Higher cognitive load and lower throughput per annotator.
Imaging Annotation
Region marking, segmentation and classification on DICOM studies, requiring both radiological expertise and tooling that handles imaging formats properly.
Structured and Signal Data Labelling
Waveform, vitals and device telemetry labelling for event detection, which requires clinical context that the raw signal alone does not carry.
Annotation Quality Assurance
An annotation dataset with unmeasured quality is not a dataset, it is an assumption. Models trained on inconsistent labels learn the inconsistency, and the failure appears at evaluation as unexplained variance rather than as an obvious data problem. Our process measures agreement continuously rather than at the end, because the correct time to discover that two annotators interpret a guideline differently is in the first hundred records, not the last thousand.
Guideline Development Before Volume
A written guideline with worked examples and explicit edge-case rules, developed on a pilot sample before the main annotation effort begins.
Overlap Sampling
A defined proportion of records annotated independently by more than one annotator, providing the basis for measuring agreement rather than estimating it.
Inter-Annotator Agreement Measurement
Agreement reported statistically, with the appropriate metric for the task type. Reported per label rather than as a single headline number, because averages hide the labels that are failing.
Adjudication and Guideline Revision
Disagreements resolved by a senior clinical reviewer, with the resolution fed back into the guideline so the same ambiguity is not relitigated repeatedly.
Gold Standard Sets
A held-out, carefully adjudicated set used to measure annotator performance over time and to catch drift as a project runs for months.
Versioning of Labels and Guidelines
Every label tied to a guideline version, so a change in definition is traceable rather than silently corrupting a dataset built over time.
Governance, IRB and Sensitive Categories
Regulatory obligations on clinical data used for AI development go well beyond HIPAA, and the additional layers apply most strongly to exactly the data categories AI teams find most valuable. Behavioral health, substance use, HIV status, genetic information and reproductive health all carry heightened protection under federal or state law, and several state health privacy statutes now reach data that HIPAA does not. Establishing which regimes apply is the first workstream on any annotation project, not a compliance review at the end.
01
When IRB Review Applies
Research involving human subjects generally requires review. Operations and quality improvement work generally does not, and the boundary is less obvious than teams assume.
02
Substance Use Records Under 42 CFR Part 2
Substance use disorder records from covered programs carry restrictions stricter than HIPAA, including on redisclosure. De-identification alone does not always resolve them.
03
Behavioral and Mental Health Data
Heightened protection under many state laws, with consent requirements that can differ from HIPAA. Treat these datasets as requiring specific legal review rather than general clearance.
04
State Health Privacy Laws
Several states now regulate health data outside HIPAA’s scope, including data held by entities that are not covered entities. Applicability depends on where the individuals are, not where you are.
05
Consent Language and Secondary Use
Whether the original consent permits model development is a question for counsel and frequently determines scope. We ask it before starting rather than after.
06
What We Will Not Do
We do not source, purchase, broker or resell clinical datasets, and we do not annotate data a client cannot demonstrate lawful basis to use for the stated purpose.
Our Clinical Data Annotation Services
We deliver annotation as scoped engagements against data you hold and control, beginning with a pilot rather than a full production run. The pilot establishes the guideline, measures achievable agreement and produces a real cost per record for your specific task, which is the only reliable basis for pricing the main effort. A Business Associate Agreement is executed before any access to protected health information, and our information security management system is ISO 27001 certified.
Feasibility and Governance Review
Assessment of lawful basis, applicable regimes, consent scope and de-identification strategy, delivered as a written position you can take to counsel.
De-identification Pipeline Build
Automated identifier removal with human review of residual risk in narrative text, plus documentation supporting whichever route you have chosen.
Annotation Schema and Guideline Design
Label definitions, edge-case rules and worked examples, developed with your clinical and machine learning stakeholders on a real sample.
Pilot Annotation and Agreement Baseline
A bounded pilot establishing achievable agreement, realistic throughput and a defensible per-record cost before committing to production volume.
Production Annotation With Continuous QA
Clinician-led labelling with overlap sampling, ongoing agreement measurement, adjudication and guideline versioning throughout.
Dataset Delivery and Documentation
Labelled data with provenance, guideline versions and agreement statistics, in a form your model pipeline consumes and your auditors accept. See our healthcare AI work.
How Annotation Work Is Priced
Annotation is priced per record for production volume and as a fixed fee for the work that precedes it, because the two have entirely different risk profiles. Nobody can quote a per-record rate honestly before the guideline exists and agreement has been measured, since the rate depends on task complexity, required annotator seniority and how much adjudication the schema generates. The pilot exists to produce that number. Any vendor quoting a production rate before a pilot is quoting a number they will revise.
Feasibility and Governance Review
A fixed-fee engagement covering lawful basis, regime applicability and de-identification strategy, delivered as written findings.
De-identification Pipeline
Priced as a build against your data types and volume, with ongoing per-record processing cost separate from the initial engineering.
Pilot Engagement
Fixed fee covering guideline development, a bounded annotation sample, agreement measurement and a production cost model. This is where most engagements start.
Production Annotation
Per-record pricing set by the pilot, varying by task type and required annotator qualification. Clinician-annotated relation extraction and simple document classification are not comparable rates.
Dedicated Annotation Team
Monthly retainer for teams running continuous annotation across multiple projects, where per-record pricing creates more administrative overhead than it removes.
What Sits Outside the Price
Cloud infrastructure, annotation tooling licences and any legal or IRB fees are quoted separately and never absorbed into a per-record rate.
FAQs
Frequently Asked Questions
Clinical data annotation is the labelling of healthcare records, notes, images or signals so they can train or evaluate a machine learning model. It covers de-identification, schema design, clinician-led labelling, agreement measurement and quality assurance.
Safe Harbor removes eighteen specified identifier categories and requires no expert sign-off. Expert Determination uses a qualified expert to document that re-identification risk is very small, which allows retaining analytically valuable fields such as dates that Safe Harbor removes.
No. We annotate data our clients already hold and control. We do not source, purchase, broker or resell clinical datasets, and we require demonstrated lawful basis before annotating anything.
It depends on whether the work constitutes human subjects research. Operations and quality improvement work generally does not require review, but the boundary is genuinely ambiguous in places and is worth establishing before annotation begins.
Through overlap sampling, where a proportion of records are annotated independently by multiple annotators, with agreement reported per label rather than as a single figure, and disagreements adjudicated back into the guideline.
No honest answer exists before a pilot. The rate depends on task complexity, required annotator seniority and adjudication volume. The pilot establishes the guideline, measures achievable agreement and produces a defensible production rate.
Send us the data types involved, approximate volume, what the model is intended to do and what consent or contractual basis covers the data. You will speak with an engineer rather than a salesperson. If the lawful basis is not clear, we will tell you that before discussing scope. Start through our contact form.