Skip to content

FUTURE-AI Alignment

DRAFT — Internal Review Only · Not for distribution in wide release

This document is a working draft intended for internal review by the Enterprise OMOP implementation team and our growing network of collaborators. It has not been finalized, peer-reviewed, or approved for external distribution.

See releases and roadmap for details.

Context

See Notes and NLP Design for the problem overview, Architecture for the full schema specification, and TRIPOD-LLM Alignment for the companion reporting guideline mapping.

From reporting to deployment

Where TRIPOD-LLM standardizes what researchers write in a paper, FUTURE-AI addresses a broader question: how should healthcare AI tools be designed, developed, validated, deployed, and monitored to be trustworthy?

Published in The BMJ in February 2025, the FUTURE-AI framework was developed by 117 interdisciplinary experts from 50 countries over a two-year modified Delphi process (Lekadir et al., 2025). It provides 30 best practices organized under six guiding principles — Fairness, Universality, Traceability, Usability, Robustness, Explainability — plus 7 cross-cutting General recommendations.

FUTURE-AI is not a publication checklist. It is a lifecycle framework that spans from initial design through post-deployment monitoring. This makes it directly relevant to how NLP infrastructure should be built — not just documented.

The six FUTURE-AI principles

Principle Core concern # of recommendations
Fairness Equal performance across patient groups; bias identification and mitigation 3
Universality Generalizability across settings, populations, and clinical workflows 4
Traceability Documentation, auditing, logging, and governance throughout the AI lifecycle 6
Usability User-centered design, human-AI interaction, oversight, training, and clinical utility evaluation 5
Robustness Resilience to real-world data variations, adversarial inputs, and distribution shifts 3
Explainability Clinically meaningful information about the logic behind AI decisions 2

Mapping FUTURE-AI to the NLP infrastructure

Fairness

FUTURE-AI Recommendation Description Infrastructure Coverage
Fairness 1: Define sources of bias Identify potential bias types and sources at design phase Publication/governance concern. Infrastructure supports post-hoc bias analysis by enabling stratified queries on _DERIVED tables joined to person demographics.
Fairness 2: Collect information on data attributes Collect demographic and data attributes to enable bias detection OMOP person table provides demographics. note_span linked to note linked to person enables demographic stratification of NLP outputs.
Fairness 3: Evaluate fairness Apply bias detection methods (statistical parity, group fairness, equalized odds) Future direction: evaluation metadata tables linking execution records to stratified performance metrics would close this gap.

Universality

FUTURE-AI Recommendation Description Infrastructure Coverage
Universality 1: Define intended clinical settings Specify populations, settings, and potential obstacles to universality Publication/governance concern.
Universality 2: Use existing standards Build on community standards (SNOMED CT, OMOP, FHIR, etc.) Core alignment. The entire architecture is built on OMOP CDM. note_span_concept maps to OMOP standard concepts. _DERIVED tables follow CDM schema. OHDSI tools work natively.
Universality 3: Evaluate using external data Validate with external datasets from multiple sites nlp_execution records from different sites can be compared. The schema is designed for cross-site federation (identified as a future direction).
Universality 4: Evaluate local clinical validity Assess performance in local clinical workflows and populations _DERIVED tables enable local validation by comparing NLP extractions against discrete EHR data at the same institution.

Infrastructure advantage

FUTURE-AI Universality 2 explicitly recommends building on OMOP. The Enterprise OMOP NLP architecture is a native OMOP extension — concept mappings, CDM-compatible _DERIVED tables, and vocabulary-driven typed extractions. This is not retrofitted alignment; the infrastructure was designed from the ground up on the standard FUTURE-AI recommends.

Traceability

The Traceability principle is the most directly aligned with the Enterprise OMOP NLP architecture. FUTURE-AI defines traceability as "mechanisms for documenting and monitoring the complete trajectory of the AI tool, from development and validation to deployment and usage."

All six Traceability recommendations map to infrastructure tables:

FUTURE-AI Recommendation Description Infrastructure Coverage
Traceability 1: Implement risk management Analyse risks throughout the AI lifecycle; maintain a risk management file nlp_system + pipeline registration provides the system-of-record for what is deployed. Risk documentation is a publication/governance concern, but the infrastructure ensures the identity of each system is unambiguous.
Traceability 2: Provide documentation Create technical documentation (model properties, hyperparameters, training data, evaluation criteria, audits) nlp_system (name, version), Component (version, data artifacts), pipeline_component (config including hyperparameters). The infrastructure is the structured documentation.
Traceability 3: Implement continuous quality control Monitor AI inputs and outputs; provide uncertainty estimates note_span.probability captures per-extraction confidence. nlp_execution records enable temporal monitoring. _DERIVED table separation enables quality comparison against discrete EHR data.
Traceability 4: Implement periodic auditing and updating Configurable system for periodic evaluation; detect data drift, concept drift, performance degradation nlp_execution records with dates enable audit trails. Multiple executions of different pipeline versions against the same notes enable before/after comparison.
Traceability 5: Implement AI logging Log data accessed, AI predictions, clinical decisions, and issues encountered note_span_execution links every output to its execution. nlp_execution captures when, which pipeline, which worker version. The full provenance chain from _DERIVED row back to note serves as the audit log.
Traceability 6: Implement AI governance Specify roles for risk management, auditing, maintenance, and supervision; assign accountability Governance is an organizational concern, but the infrastructure provides the substrate — you cannot govern what you cannot trace. The normalized hierarchy (nlp_system → pipeline → Component) makes it possible to assign ownership per system and pipeline.

Infrastructure advantage

FUTURE-AI Traceability asks teams to implement logging, auditing, documentation, and quality control. The Enterprise OMOP NLP architecture provides the schema where all of this lives — queryable, structured, and attached to every extracted fact. Without this schema, traceability recommendations become aspirational prose in a governance document.

Usability

FUTURE-AI Recommendation Description Infrastructure Coverage
Usability 1: Define user requirements Involve clinical experts and end users from early stages to gather intended use and user requirements Organizational/process concern. Infrastructure supports usability by producing outputs in standard OMOP CDM format — familiar to OHDSI researchers and compatible with existing analytic tools.
Usability 2: Establish human-AI interactions and oversight Create interfaces for effective operation and implement human-in-the-loop mechanisms for quality checks and overrides note_span.probability enables confidence-based review workflows. _DERIVED table separation from discrete EHR data supports human comparison and override. The span-level model (offsets, source text) gives reviewers the context needed to evaluate and correct extractions.
Usability 3: Provide training materials Supply tutorials, manuals, examples, and hands-on sessions Publication/documentation concern. The schema's self-describing hierarchy (nlp_system → pipeline → Component) reduces the documentation burden — the structure is the explanation.
Usability 4: Evaluate user experience Test usability in real-world clinical settings with diverse end users Organizational concern. Infrastructure supports evaluation by providing queryable execution records and outputs that can be assessed against clinical workflows.
Usability 5: Evaluate clinical utility Demonstrate benefits compared to standard care; document safety _DERIVED tables enable direct comparison of NLP-extracted data against discrete EHR data at the same institution, providing the substrate for clinical utility evaluation.

Infrastructure advantage

FUTURE-AI Usability emphasizes human-in-the-loop oversight and clinical validation. The Enterprise OMOP NLP architecture supports this through confidence scores (note_span.probability), span-level provenance for human review, and _DERIVED table separation that enables side-by-side comparison with discrete clinical data.

Robustness

FUTURE-AI Recommendation Description Infrastructure Coverage
Robustness 1: Define sources of data variation Inventory real-world variations that may affect performance Publication/governance concern. Infrastructure enables detection through temporal analysis of nlp_execution records and confidence distributions in note_span.
Robustness 2: Train with representative data Use training data that reflects real-world variation Model training concern, outside infrastructure scope.
Robustness 3: Evaluate robustness Stress tests, repeatability tests under real-world conditions Multiple nlp_execution records against the same notes with different pipeline versions enable repeatability analysis. note_span.probability distributions can be compared across executions.

Explainability

FUTURE-AI Recommendation Description Infrastructure Coverage
Explainability 1: Define explainability needs Establish whether and what type of explainability is required Publication/governance concern.
Explainability 2: Evaluate explainability Assess whether explanations are meaningful, correct, and beneficial note_span with character offsets (span_start, span_end) and span_text provides intrinsic explainability — every extraction is grounded in the exact source text. note_span_assertion captures inter-span reasoning (e.g., negation, temporal links).

Infrastructure advantage

The span-based model provides built-in explainability at the extraction level. Every NLP-derived fact in a _DERIVED table can be traced back to the exact text span, character offsets, and source note. This is not a post-hoc explanation — it is the provenance chain itself.

General

FUTURE-AI Recommendation Description Infrastructure Coverage
General 1: Engage stakeholders continuously Multidisciplinary engagement throughout the AI lifecycle Organizational concern.
General 2: Ensure data protection Privacy-enhancing techniques, data governance, access logging OMOP's de-identification model applies. nlp_execution logging supports data governance audits.
General 3: Implement measures to address AI risks Define mitigation plan aligned with application-specific requirements Risk management file is a governance artifact. Infrastructure provides the audit substrate.
General 4: Define adequate evaluation plan Test data, metrics, reference methods, benchmarking Future direction: evaluation metadata tables.
General 5: Comply with AI regulations Identify applicable regulations (EU AI Act, FDA, etc.) The infrastructure supports compliance by providing the traceability and documentation that regulations require.
General 6: Investigate ethical issues Address application-specific ethical, social, legal concerns Organizational/publication concern.
General 7: Investigate social and environmental issues Assess impact on working conditions, sustainability, societal effects Organizational/publication concern.

Summary: where infrastructure meets guidance

FUTURE-AI Principle Recommendations addressed by infrastructure Recommendations requiring governance/publication
Fairness (3) 1 of 3 — demographic stratification via OMOP person joins Bias source definition, fairness evaluation
Universality (4) 3 of 4 — OMOP standards, cross-site federation, local validation Clinical setting definition
Traceability (6) All 6 — the infrastructure is the traceability layer Risk documentation, governance roles
Usability (5) 2 of 5 — confidence-based review, span provenance for oversight, clinical utility comparison via _DERIVED tables User requirements, training materials, UX evaluation
Robustness (3) 2 of 3 — repeatability analysis, variation detection Training data representativeness
Explainability (2) 1 of 2 — span-based provenance provides intrinsic explainability Explainability needs assessment

15 of 23 principle-level recommendations (plus 2 of 7 general recommendations) are directly supported or enabled by the Enterprise OMOP NLP infrastructure. The remaining recommendations are governance, organizational, or publication concerns that require human processes — but the infrastructure provides the substrate that makes those processes auditable rather than aspirational.

FUTURE-AI and TRIPOD-LLM: complementary layers

Dimension TRIPOD-LLM FUTURE-AI Enterprise OMOP NLP Infrastructure
Scope Publication reporting Full AI lifecycle Production data infrastructure
Format Checklist for manuscripts Best practices framework Queryable database schema
Primary audience Authors and reviewers Development teams and regulators NLP engineers and data consumers
Traceability Report what was done Implement logging and auditing Store provenance at the row level
Standards — Recommends OMOP, SNOMED CT, FHIR Built natively on OMOP CDM
Lifecycle coverage Publication time Design → deployment → monitoring Every pipeline execution

All three layers are necessary. TRIPOD-LLM ensures the scientific community can evaluate the approach. FUTURE-AI ensures the development team builds responsibly. The Enterprise OMOP NLP infrastructure ensures the data itself carries the metadata that both frameworks require.

References

  • Lekadir K, Frangi AF, Porras AR, et al. FUTURE-AI: international consensus guideline for trustworthy and deployable artificial intelligence in healthcare. BMJ. 2025;388:e081554. doi:10.1136/bmj-2024-081554
  • Gallifant J, Afshar M, Ameen S, et al. The TRIPOD-LLM reporting guideline for studies using large language models. Nature Medicine. 2025;31:60–69. doi:10.1038/s41591-024-03425-5
  • Fu S, Wang L, Moon S, et al. Recommended practices and ethical considerations for natural language processing–assisted observational research: A scoping review. Clinical and Translational Science. 2023;16(3):398–411. doi:10.1111/cts.13463