Global life sciences AI · Published 16 November 2026 · HERMS Consultancy DMCC
Explore the HERMS approach and our solutions portfolio.
Preparing healthcare data for AI means demonstrating that the information is suitable for a defined use, lawfully accessible, consistently interpreted and traceable from source to output. Start with the decision and required variables—not the model. Then assess clinical meaning, missingness, timing, identity, privacy, interoperability and provenance before releasing a versioned dataset. A successful API call, de-identification step or common-data-model conversion does not independently establish reliable AI. The relevant question is whether the entire data pipeline supports the intended use and its risks. Guidance Regarding Methods for De-identification of Protected Health Information in Accordance with the HIPAA Privacy Rule FHIR Overview — FHIR v4.0.1 OMOP Common Data Model AI Risk Management Framework
AI readiness is use-specific—not a property of a data lake
A dataset can be useful for service-volume reporting but unsuitable for predicting deterioration. A billing code may describe a reimbursed activity rather than a confirmed diagnosis; a timestamp may indicate documentation rather than when an event occurred. Before cleaning anything, define the task, user, population, setting, prediction or decision time, required output and consequences of error.
Record what data will actually be available when the system is used. Specify whether the application trains a predictive model, retrieves documents for a language model or coordinates operational actions. These uses share governance needs but require different preparation and evaluation. The framework below is a proposed operating approach informed by standards and risk guidance, not a universal regulatory checklist. AI Risk Management Framework
Four foundations of healthcare AI data readiness
| Foundation | Core question | Evidence to retain |
|---|---|---|
| Quality | Is the information clinically and analytically fit for this task? | Profiles, rules, label review, missingness and temporal checks |
| Privacy and permitted use | May these data be used this way, by these parties? | Authority, roles, agreements, safeguards and retention decisions |
| Interoperability | Does meaning survive exchange and transformation? | Profiles, dictionaries, terminology mappings and interface tests |
| Provenance | Can the origin and transformation of the output be reconstructed? | Source identifiers, versions, lineage, code and release approvals |
None of the four compensates for another. Well-documented unlawful processing is still a problem. Perfectly formatted data may be clinically misleading. An accurate training dataset may fail when deployed against another hospital’s documentation practices. Treat readiness as a controlled release decision rather than a one-time cleansing exercise.
Build a data inventory and variable-level specification
Inventory sources, owners, populations, time coverage, collection purpose, permissions, refresh frequency, formats and known limitations. Include notes, attachments, images, derived fields and third-party enrichments—not only structured tables. Describe the route from the originating system to the proposed AI environment.
For each essential variable, record definition, source field, terminology, unit, valid range, timing, missing-value meaning and transformation. Distinguish prescribed medication from dispensing and administration; distinguish a problem-list entry from an incident diagnosis. Name the clinical reviewer who resolves ambiguity. A missing variable critical to the task may require changing the task or collecting different data.
Measure quality where errors affect the decision
Data quality is more than completeness. Assess conformance to the expected representation; completeness of relevant information; plausibility and consistency; and fitness for the intended clinical question. OHDSI’s Data Quality Dashboard demonstrates configurable, transparent checks against an OMOP instance. Passing a dashboard does not prove that labels, clinical definitions or the proposed AI task are valid. DataQualityDashboard
| Dimension | Example test | Why a clean-looking dataset may still fail |
|---|---|---|
| Completeness | Required fields by site, time and subgroup | Recorded values may exist only for higher-risk patients |
| Conformance | Types, codes, relationships and units | A valid code can represent the wrong clinical concept |
| Plausibility | Ranges and clinically meaningful sequences | A possible value can still belong to the wrong patient |
| Temporal integrity | Event, collection, documentation and availability times | Future information can leak into a past prediction |
| Uniqueness / identity | Duplicate events and patient/encounter linkage | Copied records inflate counts or cross evaluation splits |
| Label validity | Reference definition, adjudication and uncertainty | Billing or proxy labels may not represent the target outcome |
| Representativeness | Sites, equipment, subgroups and care settings | Large datasets can omit the intended deployment population |
Set thresholds before acceptance and match them to the use. A high overall pass rate can hide a critical unit error. Report unresolved findings and excluded records; aggressive cleaning can change the population and conceal meaningful differences. These are proposed analytical controls, not universal numerical requirements.
Missing data are part of the clinical process
A laboratory value may be absent because a test was unnecessary, unavailable, declined, performed elsewhere or not transferred. Treating all missing values as zero can invent a clinical observation. Profiling should distinguish unknown, not applicable, not measured, unavailable and explicitly negative where the source supports those distinctions.
Assess patterns by site, outcome, population and time. Choose handling methods based on the task and plausible mechanism, and document assumptions. Imputation can introduce information leakage if fitted on evaluation data. It also cannot guarantee correction of systematic under-recording or absence of a key confounder. Keep the original missingness and transformation history reviewable.
Prevent patient, time and label leakage
Define the decision time and when each predictor becomes available. A diagnosis entered after discharge may improve retrospective prediction while being unavailable at admission. Separate clinical event time from system-entry time and final-result time; preserve time zones and ordering logic.
Partition data at the appropriate patient, encounter, site and temporal level. Check repeated images, copied notes and near-duplicate records across splits. Fit preprocessing, feature selection and imputation using only the allowed development information. Independent evaluation must test the full prepared pipeline, not a model fed by transformations learned from the test set.
Make label creation explicit. Document the clinical concept, observation window, ascertainment, adjudication and unresolved disagreement. A model trained to predict a workflow artifact may reproduce documentation habits rather than the intended clinical state. These are proposed safeguards for credible evaluation within a lifecycle risk approach. AI Risk Management Framework
Privacy: removing names does not automatically anonymise data
Determine the applicable jurisdiction, institutional role and processing purpose before moving data to a model provider. Under GDPR, re-identifiable pseudonymised data remains personal data. Processing health or genetic information requires assessment of the relevant lawful basis and special-category condition, with national requirements where applicable. Research is not a blanket exemption, and consent is not the only possible route. Data protection explained Legal grounds for processing data
In the US HIPAA context, HHS describes Safe Harbor and Expert Determination as two methods for satisfying the Privacy Rule de-identification standard. Safe Harbor is not merely deleting names: specified identifiers and the actual-knowledge condition matter, including in free text. HHS also notes that properly de-identified information retains a very small, non-zero identification risk. HIPAA scope and GDPR anonymisation should not be treated as interchangeable. Guidance Regarding Methods for De-identification of Protected Health Information in Accordance with the HIPAA Privacy Rule
Assess rare conditions, genomic information, exact dates, images, narrative notes and linked datasets. Removing an obvious identifier can leave a distinctive combination. Synthetic data, encryption, federated learning and a secure enclave may support safeguards, but none automatically proves anonymity or lawful processing. Evaluate the actual design and residual risk.
Control the entire data flow, including prompts and logs
Use least-privilege access, appropriate authentication, protected transfer and storage, approved environments and reviewable access history. Map subprocessors, hosting, remote administration, provider retention and any use of submitted material for model training. Check local requirements before cross-border transfers or access; selecting a hosting region alone does not settle every legal question.
Minimise data in prompts, debugging output, embeddings, caches, exports and support tickets. Treat derived stores as governed assets and assess what sensitive information they expose. Define retention and deletion responsibilities, including backups and derived copies, while preserving records that must lawfully be retained. Security arrangements support privacy; they do not replace authority to process. Guidance Regarding Methods for De-identification of Protected Health Information in Accordance with the HIPAA Privacy Rule Data protection explained Legal grounds for processing data
Interoperability requires shared meaning—not just an API
Technical connectivity is the ability to transfer information. Structural consistency makes fields understandable to software. Semantic interoperability requires consistent clinical meaning. A JSON response can be syntactically valid while mixing units, confusing historical and current diagnoses or dropping negation.
HL7 FHIR supports resource-based electronic exchange. Profiles, terminology bindings and implementation guides constrain how it is used. Specify the required FHIR release and actual implementation guide; an undifferentiated “FHIR-compatible” claim is insufficient. The R4 pages cited here are stable version-specific references, not a claim that R4 is the newest release or universally mandatory. FHIR Overview — FHIR v4.0.1
OMOP CDM supports standardized structure and content for observational analytics. FHIR and OMOP have different emphases and may complement one another; neither automatically replaces all source systems or guarantees correct mapping. Preserve source codes, vocabulary versions and mapping decisions, and assess loss of detail during transformation. FHIR Overview — FHIR v4.0.1 OMOP Common Data Model
| Representation issue | Potential downstream error | Proposed check |
|---|---|---|
| Units and reference ranges | Magnitude interpreted incorrectly | Validate analyte, unit, specimen and conversion together |
| Terminology mapping | Different concepts collapsed into one label | Clinician review; unmapped-code monitoring; versioned mappings |
| Medication representation | Order mistaken for administration | Retain event type, route, dose and status |
| Identity linkage | Another patient’s event is included | Assess matching quality and uncertain links |
| Time representation | Events appear in the wrong clinical order | Track event, entry and availability times with zones |
| Narrative context | Historical/negated condition becomes a current diagnosis | Test negation, experiencer, temporality and language |
Provenance: preserve the source-to-output chain
Provenance explains who or what created information, which sources were used and which activities changed it. A practical lineage chain links source extraction, mapping, cleaning, labeling, cohort selection, feature creation, dataset release and output. Keep source identifiers and versions, processing code, configurations, timestamps, responsible actors and material review decisions.
FHIR Provenance can describe activities creating, revising or signing a resource version and the involved entities and agents. AuditEvent addresses overlapping but different event-auditing needs. Neither simply replaces the other. In R4, Provenance is a trial-use resource; implementing it is not in itself proof of legal signature compliance or trustworthy clinical content. Resource Provenance — FHIR v4.0.1
Provenance is also not indiscriminate retention of patient material. Retain decision-relevant metadata and secure references with suitable access controls. A cryptographic hash may help detect content change but does not establish clinical correctness, lawful collection or reviewer competence. Test whether a reviewer can reconstruct a sampled output without granting unnecessary access.
Prepare retrieval data differently from predictive-model data
For retrieval-augmented generation, define approved source types, owners, scope, version, effective date and retirement rules. Check document extraction, OCR, tables, section boundaries and permission propagation. Preserve citation anchors so a generated statement can be checked against the supporting passage.
Do not mix obsolete and current protocols without a resolution policy. Prevent retrieved content from being treated as executable instructions. Evaluate retrieval coverage and source-to-claim fidelity separately from fluent answer quality. Deleting a superseded document from the source repository is not enough if the search index or cache still serves it.
For predictive work, focus additionally on decision-time availability, label validity and split independence. For agents, add action permissions and accountable approval boundaries. One data lake can support all three, but each use needs a documented preparation and evaluation contract. AI Risk Management Framework
Common reasons healthcare AI outputs become unreliable
| Failure | Why it can be missed | Practical response |
|---|---|---|
| Wrong units or field mapping | Values look numeric and complete | Clinically verify meaning and boundary conversions |
| Future-information leakage | Retrospective accuracy appears excellent | Audit feature availability at decision time |
| Selection and recording bias | Large sample masks missing care settings | Profile population and capture mechanisms |
| Poor labels | Proxy appears objective | Validate outcome definition and adjudication |
| Missingness encoded as normal | Pipeline accepts a default value | Preserve missingness semantics and test handling |
| Stale source documents | Retrieval returns confident citations | Version control, retirement and cache checks |
| Untraceable transformations | Output exists but cannot be explained | Versioned lineage and reproducible release packages |
| Dataset drift | Training checks are never repeated | Monitor changes and reassess readiness |
| Permission leakage | Data appear in logs or retrieval beyond intended users | End-to-end access testing and minimisation |
A practical healthcare AI data-readiness checklist
Use this proposed gate checklist for a defined use case. No weighted score should override an unresolved unlawful-use question or a critical integrity defect.
- Purpose: intended users, decision, population and consequences are documented.
- Authority: permitted use, institutional roles, agreements and applicable review are confirmed.
- Inventory: sources, dates, owners, coverage and restrictions are recorded.
- Meaning: essential variables, units, terminology and timing are defined.
- Quality: risk-based checks and acceptance thresholds are approved; critical findings resolved.
- Labels: target definitions, reference standards and uncertainty are reviewable.
- Independence: leakage and duplication across development/evaluation boundaries are assessed.
- Privacy/security: re-identification and access risks cover source, derived stores, providers and logs.
- Interoperability: profiles, releases and mappings are versioned and clinically checked.
- Provenance: a sampled result can be traced to sources and transformation versions.
- Release: dataset, code, configuration and accountable approval are bound together.
- Lifecycle: refresh, monitoring, incidents, retirement and reassessment triggers are assigned.
Hypothetical example: combining hospital laboratory datasets
A research team plans a model using laboratory data from several hospitals. Schema checks pass, but clinical review finds incompatible units, site-specific missing-value codes and timestamps that sometimes reflect verification rather than collection. Notes also contain residual identifiers.
The team pauses release, documents source-specific mappings, validates conversions, preserves missingness and timing semantics, and routes privacy questions to the responsible specialists. It revisits the prediction-time definition and rebuilds independent evaluation splits. The release package records the source and transformation versions, unresolved limitations and approval.
This example illustrates a readiness workflow, not a validated HERMS deployment or an asserted performance gain. The lesson is that the model was not the first problem: apparently interoperable input concealed differences in clinical meaning.
Legal frameworks and standards: know what applies
| Resource | Status / scope | Practical implication |
|---|---|---|
| GDPR and national laws | Binding where applicable; Commission pages are explanatory | Assess personal/special-category data, roles, purpose and safeguards |
| HIPAA Privacy Rule | Binding for relevant US entities and PHI; HHS guidance explains methods | De-identification requires an applicable method, not a marketing label |
| EHDS | EU regulation with phased application | Do not infer immediate access or a general AI-training permission |
| HL7 FHIR | Versioned interoperability specification | Specify profiles/releases; transport does not prove clinical meaning |
| OMOP / OHDSI | Community data standard and tools | Assess transformation quality and compatibility; not legal certification |
| NIST AI RMF | Voluntary cross-sector framework; revision underway | Organize lifecycle risk assessment without claiming regulatory approval |
The Commission states that EHDS entered into force on 26 March 2025, with key requirements applying in phases in 2029 and 2031. As of 16 November 2026, entry into force should not be confused with full operational application. EHDS builds on existing frameworks including GDPR; it does not mean all health data are freely accessible for AI development. Confirm the applicable provisions and national arrangements for a specific project. European Health Data Space Regulation (EHDS)
Implementation roadmap and measurable readiness
| Planning window | Proposed output | Accountable lead |
|---|---|---|
| Days 1–30 | Use-case/data inventory; authority and flow map; initial quality profile | Process and data owners with privacy/security |
| Days 31–60 | Variable specifications, mappings, lineage and reproducible pipeline; critical remediation | Clinical data lead and engineering |
| Days 61–90 | Independent review, acceptance evidence, controlled release and monitoring | Quality/assurance and use-case owner |
These windows are illustrative, not delivery promises. Track essential-field validity by site; clinically verified mapping coverage; provenance completeness; unresolved critical defects; inappropriate-access findings; reproducible-release success; and time to investigate a detected change. Define denominators and owners before reporting. Avoid a single “AI-ready” percentage that hides critical gaps.
HERMS perspective: make data decisions reviewable
HERMS positions ATTESTRA around integrity and data governance and VERIQIS around evidence integrity and assurance. These are company descriptions, not independently validated integrations, privacy certifications or performance claims. Relevant evaluation should test whether the proposed workflow can preserve source meaning, permissions, traceability and accountable approval for the intended setting. AI Life Sciences Software | 25 HERMS Solutions
Explore HERMS integrity, data-governance and evidence solutions to discuss a demonstration focused on your data-readiness gaps and evidence requirements.
Frequently asked questions
What makes healthcare data AI-ready?
Fitness for a defined task, lawful and controlled access, meaningful variables, appropriate quality, traceable transformations and tested lifecycle controls. Readiness is conditional on the use and environment. A dataset that supports reporting may not support a clinical model or an externally shared research analysis.
Is removing names enough to anonymise health data?
No. Dates, rare conditions, images, narrative text and combinations of attributes may permit identification. HHS specifies two HIPAA de-identification methods; GDPR separately treats re-identifiable pseudonymised data as personal data. Assess the actual legal context and risk rather than assuming one global test. Guidance Regarding Methods for De-identification of Protected Health Information in Accordance with the HIPAA Privacy Rule Data protection explained
Does FHIR guarantee interoperable AI data?
No. FHIR supports structured exchange, but implementations need the appropriate version, profiles, terminology and testing. A valid resource can still contain the wrong clinical meaning, unit, patient reference or time interpretation. Evaluate semantic correctness and task suitability in addition to technical conformance. FHIR Overview — FHIR v4.0.1
Are FHIR and OMOP the same?
No. FHIR primarily supports interoperable information exchange; OMOP standardizes observational data for analytics. They can complement one another. Mapping between them requires explicit clinical and technical choices and may lose information; neither conversion guarantees privacy or fitness for a study. FHIR Overview — FHIR v4.0.1 OMOP Common Data Model
What is data provenance in healthcare AI?
It is the reviewable record of origins, actors and transformations that produced a dataset or output. Preserve relevant source versions, processing configurations, mappings and release decisions. Provenance helps assess reliability; it does not itself prove clinical accuracy or lawful processing. Resource Provenance — FHIR v4.0.1
Can synthetic or federated data solve privacy concerns?
Not automatically. Assess the method, what can be inferred or disclosed, the operating environment and applicable legal duties. Synthetic outputs may preserve sensitive patterns, and federated systems still need access, security and governance controls. These approaches should be assessed as safeguards, not universal exemptions.
Why can a high-performing model fail in another hospital?
Differences in case mix, recording practices, equipment, units, missingness, care pathways and timing can change what the inputs mean. Strong retrospective metrics do not establish performance in every deployment. Assess local compatibility and representative evaluation, and investigate changes after release.
How often should data readiness be reassessed?
At risk-appropriate intervals and when material changes occur: source systems, schemas, terminology, populations, permissions, retrieval content or intended use. Assign triggers and owners. A previous release decision should not silently authorize every future refresh or a different downstream application. AI Risk Management Framework
Reliable healthcare AI starts with a defensible data decision. Before asking whether the model is sophisticated enough, ask whether the information is suitable, permitted, understood and traceable. If those conditions cannot be demonstrated, pause, narrow the use or improve the evidence.
Discuss a data-readiness or governance-focused demonstration through HERMS Solutions , or explore related HERMS healthcare and life-sciences insights .
Sources and further reading
- Guidance Regarding Methods for De-identification of Protected Health Information in Accordance with the HIPAA Privacy Rule
- FHIR Overview — FHIR v4.0.1
- OMOP Common Data Model
- AI Risk Management Framework
- DataQualityDashboard
- Data protection explained
- Legal grounds for processing data
- Resource Provenance — FHIR v4.0.1
- European Health Data Space Regulation (EHDS)
Explore the pharmaceutical AI series
- AI decision authority and human oversight
- Agentic versus generative AI in pharmaceuticals
- AI governance in life sciences
- Human approval controls for pharmaceutical AI
- Real-world evidence for regulatory and market access
- AI medical device validation and change control
- Global pharmaceutical market access, HTA and pricing evidence
Explore HERMS life sciences solutions and AI governance and traceability, or request a workflow-focused demonstration.
Relevant HERMS solutions
Explore the HERMS solutions relevant to this article and assess their fit for your workflow, evidence and approval requirements.
