FAIR Data Principles for AI: Making Your Training Data Reusable

FAIR data principles provide a practical framework for preparing AI training datasets in bioinformatics so that they remain findable, accessible, interoperable, and reusable by other researchers and automated systems. For students, researchers, analysts, and life-science professionals building machine learning models, applying FAIR principles at the dataset creation stage reduces duplicated effort, supports regulatory compliance, and improves the reproducibility of model outcomes. This article explains how to implement each FAIR component for training data, with concrete steps for metadata creation, provenance tracking, licensing decisions, and quality control, using examples from genomics, proteomics, imaging, and clinical informatics.

At a Glance

The table below summarizes the four FAIR pillars and their practical application to AI training datasets in bioinformatics.

The FAIR Guiding Principles were published in Scientific Data and define these four components as the foundation for data sharing in research [4]. For AI training data, reusability depends on the other three components working together, because a model trained on poorly documented data cannot be validated or compared with other work.

Why FAIR Data Matter for AI Training in Bioinformatics

AI models in bioinformatics depend on large, diverse, and well-annotated datasets. Most laboratories cannot acquire such datasets on their own, which makes data sharing and open science initiatives essential for training effective models [8]. However, shared data only help if they arrive in a usable form. The FAIR principles state general requirements for useful data sharing, and these requirements become more stringent when data feed machine learning pipelines [8].

The consequences of non-FAIR training data appear throughout the model lifecycle. In molecular informatics, most data used for training and testing deep learning models are not available as FAIR and open data, which constrains progress in areas like synthetic chemistry and drug discovery [12]. In clinical informatics, reproducibility of machine learning applications relies heavily on data preparation, yet preprocessing pipelines are not systematically reported in a consistent way, limiting adaptability when code and datasets are shared [15]. These gaps mean that even well-executed models cannot be reproduced, compared, or extended by other groups.

FAIR data also support the ethical deployment of AI in medicine. Ethics in medical AI involves ensuring that technologies are safe, fair, and respect patient privacy, including concerns about diagnostic accuracy, fairness in patient treatment, and protection of personal health data [5]. Transparent data practices give reviewers and regulators the information they need to assess whether a model was trained on appropriate data. Bias in machine learning models can arise from training data, algorithmic choices, feature engineering, reporting practices, and temporal changes in technology or disease patterns [6]. Documenting these aspects through FAIR metadata makes bias sources visible and addressable.

The Four FAIR Principles Applied to Training Data

Findable

A findable dataset has a globally unique and persistent identifier, rich metadata, and registration in a searchable resource. For bioinformatics training data, this means assigning accession numbers through established repositories and describing the dataset with enough detail that other researchers can locate it.

The National Center for Biotechnology Information provides data resources that support findability for genomics and related fields [2]. When you deposit training data in NCBI repositories, you receive stable identifiers that can be cited in publications and model documentation. The European Bioinformatics Institute offers training and resources that emphasize the importance of structured data deposition [1].

Practical steps for findability include:

• Register the dataset in a recognized repository appropriate to the data type
• Assign a persistent identifier such as a DOI or accession number
• Create metadata that describes the biological context, experimental methods, and data collection conditions
• Include information about the version of the dataset and any updates
• Link the dataset to related publications, protocols, and analysis code

For AI training specifically, metadata should also describe how the data were split into training, validation, and test sets. This information helps downstream users understand whether a model was evaluated on independent data.

Accessible

Accessibility means that data can be retrieved using their identifier, through a standardized protocol. Accessibility does not necessarily mean open access, because human data often fall under special legal protection [8]. The key requirement is that the conditions for access are clear and machine-readable.

In bioinformatics, accessibility ranges from fully open downloads to controlled access through data access committees. The National Institutes of Health Genomic Data Sharing Policy describes expectations for sharing genomic data generated with NIH funding, including how to manage access to sensitive data [3]. Researchers building AI training datasets must determine which access model applies to their data and document it clearly.

For training data, accessibility considerations include:

• Choosing a repository that supports the required access model
• Documenting authentication and authorization procedures
• Providing clear instructions for requesting access to controlled datasets
• Ensuring that data remain accessible over time through repository commitments
• Including contact information for data access questions

Federated learning offers an alternative accessibility model. Instead of moving data to a central location, federated learning leverages data across institutions while complying with data-sharing restrictions and protecting patient privacy [22]. This approach allows AI training on distributed datasets that cannot be openly shared, which is particularly relevant for clinical bioinformatics.

Interoperable

Interoperability requires that data use community standards, controlled vocabularies, and formats that allow integration with other datasets and analysis tools. For AI training data, interoperability determines whether a dataset can be combined with other datasets or processed by standard pipelines.

The Harvard Electroencephalography Database provides an example of interoperability in practice. This resource aggregates more than 280,000 EEG recordings from more than 108,000 patients across four hospitals, with data harmonized using the Brain Imaging Data Structure and hosted on the Brain Data Science Platform [11]. The use of a standard data structure makes the dataset usable for AI-driven research and reproducible analysis [11].

Interoperability strategies for training data include:

• Using community-standard file formats for the data type
• Mapping variables to standard ontologies and controlled vocabularies
• Adopting common data models where they exist
• Documenting the versions of standards and software used
• Providing data dictionaries that define every field

The Observational Medical Outcomes Partnership Common Data Model and its Medical Imaging extension demonstrate how standardized data models support interoperable AI research. In a case replication study in Alzheimer's disease, clinical and imaging data were harmonized within this framework, with MRI acquisition metadata extracted from DICOM and mapped to standardized concepts [13]. This approach enabled reproducible image processing and interoperable reuse of machine learning workflows [13].

Reusable

Reusability is the ultimate goal of FAIR data and the component most directly relevant to AI training. A reusable dataset carries detailed provenance, licensing, and quality information that permit appropriate reuse by others.

Provenance for training data includes the complete history of how the data were collected, processed, and prepared for model training. In proteomics, reuse and translational application of datasets are limited by inconsistent standards, insufficient metadata, and inadequate computational interoperability [16]. The FAIR4prep framework addresses this gap for clinical informatics by defining minimum actionable principles for reporting preprocessing steps, with a JSON-LD schema that makes preprocessing documentation machine-readable [15].

Reusability requirements for training data include:

• Complete provenance records from raw data to final training files
• Clear licensing terms that specify permitted uses
• Documentation of ethical approvals and consent for data use
• Description of data quality checks and known limitations
• Information about preprocessing steps, including software versions and parameters
• Contact information for questions about data reuse

Legal considerations affect reusability. Copyright and fair use in AI training data present legal challenges that vary by jurisdiction [24]. Copyright protection during the training stage of generative AI differs between U.S. law, which is industry-oriented, and EU law, which is rights-oriented, with ongoing discussion of fair remuneration rights under international governance regimes [25]. Researchers assembling training datasets must understand the legal status of their data sources and document the permissions they hold.

Building a FAIR Training Dataset

Step 1: Plan for FAIR from the Start

FAIR practices should be adopted during data collection instead of added after the fact [16]. Planning for FAIR from the start reduces the effort required to make data reusable and avoids the common problem of reconstructing provenance from incomplete records.

The planning phase should address:

• Which repository will host the data
• Which metadata standard applies to the data type
• Which file formats will be used for raw and processed data
• Which access model applies to the data
• Which licenses are compatible with the intended uses
• Which ethical and legal approvals are required

For human data, legal frameworks regulating how and what data can be openly shared differ greatly across countries, which can complicate data sharing or discourage researchers from doing so [8]. Early planning should include consultation with institutional review boards and data protection officers.

Step 2: Collect Data with Metadata in Mind

Metadata collection should happen alongside data collection, not after. For bioinformatics training data, metadata typically includes:

• Experimental design and conditions
• Sample characteristics and demographics
• Instrument and protocol details
• Quality control measurements
• Dates and locations of data collection
• Personnel responsible for data collection

The Harvard Electroencephalography Database illustrates comprehensive metadata practices. EEG data are linked with clinical notes, ICD-10 codes, medications, and EEG reports, with findings extracted using expert curation, regular expressions, and medical natural language processing models [11]. This rich metadata supports multimodal analysis and improves the utility of the dataset for AI training [11].

Step 3: Process Data with Reproducible Pipelines

Data preprocessing for AI training should be documented with the same rigor as the model itself. The FAIR4prep framework encourages transparent, consistent, FAIR reporting of data preparation steps, making pipelines easier to understand, reuse, and compare across studies [15].

Reproducible preprocessing requires:

• Version-controlled code for all processing steps
• Documented software versions and parameters
• Containerized environments where feasible
• Logs that record inputs and outputs at each step
• Random seeds for any stochastic operations

Computational reproducibility in cognitive neuroscience research depends on complex linked operations that combine multiple data types, specialized toolboxes, custom scripts, and changing computing environments [17]. Open science practices and FAIR principles improve how research objects are shared, but they do not guarantee that the same analysis can be rerun to reconstruct the original result [17]. Reproducible preprocessing pipelines close this gap.

Step 4: Document Provenance and Licensing

Provenance documentation for training data should allow a downstream user to trace every training example back to its source. This includes:

• Raw data files and their identifiers
• Processing scripts and their versions
• Software environments and dependencies
• Manual curation steps and who performed them
• Quality filters and exclusion criteria
• Data splits and how they were generated

Licensing decisions should be made explicitly and documented in the dataset metadata. The legal landscape for AI training data is evolving, with different approaches in different jurisdictions [24][25]. Researchers should choose licenses that reflect their rights to the data and their intentions for reuse.

Step 5: Deposit in an Appropriate Repository

Repository selection depends on the data type, access requirements, and community norms. NCBI provides data resources for genomics and related fields [2], while other domain-specific repositories exist for imaging, proteomics, and clinical data. The European Bioinformatics Institute offers training and resources for data deposition [1].

Repository considerations include:

• Whether the repository assigns persistent identifiers
• Whether the repository supports the required access model
• Whether the repository has commitments to long-term preservation
• Whether the repository accepts the relevant metadata standards
• Whether the repository integrates with other research infrastructure

Step 6: Validate FAIR Compliance

Before releasing a training dataset, validate that each FAIR component has been addressed. This validation should be documented and included with the dataset.

A practical validation checklist includes:

• Does the dataset have a persistent identifier?
• Is the metadata complete and machine-readable?
• Can the data be accessed through a standardized protocol?
• Are access conditions clearly documented?
• Do the data use community standards and controlled vocabularies?
• Is provenance complete from raw data to final files?
• Are licensing terms explicit and machine-readable?
• Are known limitations and quality issues documented?

Metadata Standards for AI Training Data

Metadata standards vary by data type and community. For AI training data, metadata must support both human understanding and machine processing.

Genomics and Sequencing Data

Genomics data typically use standards developed by the genomics community, with repositories at NCBI [2] and EMBL-EBI [1]. Metadata for sequencing data includes library preparation methods, sequencing platform, read length, coverage, and reference genome version.

Imaging Data

The Brain Imaging Data Structure provides a standard for neuroimaging data organization and metadata. The Harvard Electroencephalography Database uses this structure to harmonize EEG data from multiple hospitals [11]. For medical imaging more broadly, the OMOP Medical Imaging extension provides a framework for standardizing imaging metadata and derived measurements [13].

Clinical Data

Clinical data require careful attention to privacy and interoperability. The OMOP Common Data Model provides a standardized approach for clinical variables, enabling reproducible cohort definition and analysis [13]. Clinical metadata must include information about data collection sites, coding systems, and any transformations applied.

Proteomics and Molecular Data

Proteomics data require standards for mass spectrometry parameters, protein identification methods, and quantification approaches. FAIR practices in proteomics should be adopted during data collection to enable reproducible and interpretable modeling [16]. Metadata should capture instrument settings, acquisition modes, and data processing parameters.

Provenance Tracking for Training Data

Provenance tracking for AI training data extends beyond the data itself to include the decisions made during dataset construction. These decisions affect model performance and should be transparent to downstream users.

Data Selection Decisions

Document which data were included and excluded from the training set, and why. Exclusion criteria might include:

• Data quality failures
• Missing critical metadata
• Duplicate samples
• Ethical or legal restrictions
• Class imbalance considerations

Data Transformation Decisions

Document all transformations applied to the raw data, including:

• Normalization and scaling methods
• Imputation strategies for missing values
• Augmentation techniques
• Feature engineering steps
• Dimensionality reduction methods

Data Split Decisions

Document how data were divided into training, validation, and test sets. This includes:

• The splitting method and random seed
• Whether splits account for data leakage risks
• Whether samples from the same patient or experiment are kept together
• Whether the test set reflects the target population

The Federated Tumor Segmentation Challenge demonstrates the importance of evaluating AI algorithms across diverse data sources. The challenge benchmarked segmentation algorithms on data distributed internationally across 32 institutions, revealing that good average performance can mask data-specific modes of failure [10]. Provenance documentation that captures data source diversity helps users understand when a model might fail.

Licensing and Legal Considerations

Licensing is a critical component of reusability for AI training data. Without clear licensing terms, potential users cannot know whether they have permission to use the data for model training.

Types of Data Licenses

Data licenses range from public domain dedications to restrictive agreements that limit use to specific purposes. Common approaches include:

• Creative Commons licenses with various conditions
• Open Data Commons licenses
• Custom data use agreements for controlled access data
• Material transfer agreements for biological materials

Copyright and Training Data

The legal status of using copyrighted data for AI training remains unsettled in many jurisdictions. Copyright protection during the training stage of generative AI differs between the U.S. and EU approaches, with ongoing international discussions about fair remuneration rights [25]. Researchers should stay informed about legal developments and document their reasoning for data use decisions.

Human Data and Privacy

Human data require additional legal and ethical considerations. The legal frameworks regulating how and what data can be openly shared differ greatly across countries [8]. The NIH Genomic Data Sharing Policy describes expectations for genomic data generated with NIH funding, including privacy protections and access management [3].

For human neuroscience data, legal frameworks and missing metadata standards present dual challenges [8]. Researchers must navigate privacy regulations while also ensuring that data are annotated with sufficient metadata to be interpretable and useful [8].

Quality Control for Training Data

Quality control for AI training data should address both data quality and dataset-level properties that affect model performance.

Data-Level Quality Checks

Individual data quality checks depend on the data type but commonly include:

• Technical quality metrics from the instrument or platform
• Outlier detection and review
• Consistency checks across related fields
• Validation against known standards or controls
• Visual inspection for imaging data

Dataset-Level Quality Checks

Dataset-level checks address properties that affect model training:

• Class balance across outcome categories
• Distribution of key variables compared with the target population
• Duplicate or near-duplicate samples
• Temporal or geographic biases
• Batch effects from different collection sites or times

The systematic review of spatio-temporal COVID-19 modeling found that geographic coverage was uneven, with a strong concentration of studies in high-income regions and persistent underrepresentation of low- and middle-income contexts [18]. Similar representativeness issues affect AI training datasets and should be documented as limitations.

Bias Assessment

Bias in machine learning models can be categorized into data bias, development bias, and interaction bias [6]. Data bias arises from the training data itself, including issues with feature engineering, reporting, and temporal changes [6]. Assessing training data for potential bias sources should be part of the dataset release process.

For medical AI, fairness in patient treatment is a core ethical concern [5]. Training data that underrepresent certain patient groups can lead to models that perform poorly for those groups. Documenting the demographic and clinical characteristics of training data supports transparency about model limitations.

Common Failure Patterns in FAIR Data Implementation

Metadata as an Afterthought

Many datasets are collected and processed before metadata documentation begins. This approach leads to incomplete or inaccurate metadata because the details of data collection decisions have been lost. FAIR practices should be adopted during data collection [16].

Repository Selection Delays

Choosing a repository after data collection can create problems if the repository has specific format or metadata requirements. Repository selection should happen during the planning phase.

Overly Restrictive or Vague Licensing

Some datasets are released without clear licensing terms, making them unusable for AI training because potential users cannot verify their rights. Other datasets use licenses that are more restrictive than necessary, limiting legitimate reuse.

Incomplete Provenance

Provenance records that capture only the final processing steps miss the decisions made during data collection and initial processing. Complete provenance requires documentation at every stage.

Ignoring Interoperability Standards

Datasets that use idiosyncratic formats or custom vocabularies cannot be integrated with other datasets or processed by standard tools. Interoperability requires adopting community standards even when they require additional effort.

Privacy Overcorrection

Some researchers respond to privacy concerns by making data inaccessible or by removing so much detail that the data lose analytical value. The legal frameworks for human data require adaptation of procedures instead of avoidance of sharing [8]. Deidentification standards, such as the HIPAA Safe Harbor approach used by the Harvard Electroencephalography Database, can support privacy-compliant data sharing [11].

Records and Measurements for FAIR Compliance

What to Record

Maintain records that document FAIR implementation decisions:

• Dataset identifier and repository information
• Metadata schema and version used
• File formats and their versions
• Software tools and versions used for processing
• Data access procedures and contact information
• License terms and any restrictions
• Ethical approvals and consent documentation
• Quality control results and known limitations
• Version history for the dataset

How to Measure FAIR Compliance

FAIR compliance can be assessed through structured evaluation:

• Check that each dataset has a resolvable persistent identifier
• Verify that metadata are available in a machine-readable format
• Confirm that access conditions are documented and testable
• Review file formats against community standards
• Assess provenance completeness from raw data to final files
• Verify that license terms are explicit and machine-readable

Automated FAIR assessment tools exist, but manual review by domain experts remains important for evaluating whether metadata adequately describe the biological and technical context.

Limitations and Professional Escalation Criteria

Limitations of FAIR Data for AI

FAIR data principles address data sharing and reuse but do not guarantee that data are suitable for a particular AI application. Additional considerations include:

• Whether the data represent the target population for the model
• Whether the data quality is sufficient for the intended analysis
• Whether the data contain biases that could lead to unfair outcomes
• Whether the data are current or reflect outdated practices

FAIR principles also do not address the computational reproducibility of analysis workflows. Open science practices and FAIR principles improve how research objects are shared, but they do not guarantee that the same analysis can be rerun to reconstruct the original result [17]. Reproducibility requires additional practices from software engineering and machine learning operations [17].

When to Seek Professional Guidance

Escalate to institutional data governance, legal, or ethics professionals when:

• The data include human subjects with complex consent or privacy requirements
• The data come from multiple jurisdictions with conflicting legal frameworks
• The intended use raises copyright questions that are not settled
• The data include sensitive information beyond standard health data
• The dataset will be shared across institutional or national boundaries
• The training data will be used for clinical decision support or other high-stakes applications

The relationship between ethics and AI in medicine requires involvement of healthcare professionals, AI experts, patients, and ethics specialists to guide and oversee implementation [5]. Researchers building training datasets for medical AI should engage these stakeholders early in the process.

Frequently Asked Questions

What is the difference between FAIR data and open data?

FAIR data are findable, accessible, interoperable, and reusable, but accessibility does not mean open access. Human data can fall under special legal protection, and FAIR principles accommodate controlled access models where authentication and authorization are documented [8]. Open data are a subset of FAIR data where access is unrestricted. Many bioinformatics datasets are FAIR but not open because they contain human data with privacy requirements.

How do FAIR principles apply to data that cannot be shared openly?

FAIR principles require that data be accessible through standardized protocols, but the access conditions can include authentication and authorization. For controlled access data, the repository should document the request process, the criteria for approval, and the terms of use. Federated learning provides another approach, allowing AI training across institutions while complying with data-sharing restrictions and protecting patient privacy [22].

What metadata are essential for AI training datasets?

Essential metadata for AI training datasets include the experimental design, sample characteristics, data collection methods, processing steps, quality control results, data splits, and known limitations. The FAIR4prep framework emphasizes that preprocessing pipelines should be reported in a consistent, standardized way to support reproducibility [15]. For clinical data, metadata should also include coding systems and any transformations applied.

How do I choose a repository for my training data?

Repository selection depends on the data type, access requirements, and community norms. NCBI provides data resources for genomics and related fields [2], and EMBL-EBI offers training and resources for data deposition [1]. Consider whether the repository assigns persistent identifiers, supports the required access model, has long-term preservation commitments, and accepts the relevant metadata standards.

What licensing considerations apply to AI training data?

Licensing for AI training data must address the rights of the data creators, any underlying data sources, and the intended uses by downstream users. Copyright protection during the training stage of generative AI differs across jurisdictions, with the U.S. and EU taking different approaches [25]. Researchers should document their rights to the data and choose licenses that reflect their intentions for reuse.

How can I assess whether my training data are biased?

Bias assessment should examine the training data for representativeness relative to the target population, including demographic, geographic, and temporal coverage. Bias in machine learning models can arise from training data, algorithmic choices, feature engineering, reporting practices, and temporal changes in technology or disease patterns [6]. Documenting the characteristics of training data supports transparency about model limitations.

What is the role of data standards in AI training data interoperability?

Data standards enable different datasets to be combined and processed by common tools. The Harvard Electroencephalography Database uses the Brain Imaging Data Structure to harmonize EEG data from multiple hospitals [11]. The OMOP Common Data Model and its Medical Imaging extension provide a framework for standardizing clinical and imaging data [13]. Adopting community standards is essential for interoperability.

How do FAIR principles support ethical AI development?

FAIR principles support ethical AI by making training data transparent and auditable. Ethics in medical AI involves ensuring that technologies are safe, fair, and respect patient privacy [5]. Documenting data provenance, limitations, and access conditions allows reviewers to assess whether a model was trained on appropriate data and whether bias sources are visible [6].

Related Bioinformatics Guides

• Ethical Considerations in Computational Genomics
• The Protein Data Bank (PDB): Archival Standards, Structural Validation Metrics, and Bioinformatics Integration Protocols
• The History and Evolution of Bioinformatics
• Bioinformatics in Agriculture and Crop Improvement
• ChIP-Seq Bioinformatics Workflows

References and Further Reading

• EMBL-EBI Training. European Bioinformatics Institute.
• NCBI Data Resources. National Center for Biotechnology Information.
• Genomic Data Sharing Policy. National Institutes of Health.
• The FAIR Guiding Principles. Scientific Data.
• Ethics and artificial intelligence.. Revista clinica espanola, 2024.
• Ethical and Bias Considerations in Artificial Intelligence/Machine Learning.. Modern pathology : an official journal of the United States and Canadian Academy of Pathology, Inc, 2025.
• AI firms must play fair when they use academic data in training.. Nature, 2024.
• FAIR human neuroscientific data sharing to advance AI driven research and applications: Legal frameworks and missing metadata standards.. Frontiers in genetics, 2023.
• Embryo selection through artificial intelligence versus embryologists: a systematic review.. Human reproduction open, 2023.
• Towards fair decentralized benchmarking of healthcare AI algorithms with the Federated Tumor Segmentation (FeTS) challenge.. Nature communications, 2025.
• Harvard Electroencephalography Database: A comprehensive clinical electroencephalographic resource from four Boston hospitals.. Epilepsia, 2025.
• Open data and algorithms for open science in AI-driven molecular informatics.. Current opinion in structural biology, 2023.
• Advancing the FAIRness of Multimodal Imaging Research Through the OMOP MI-CDM Framework: A Case Replication Study in Alzheimer's Disease.. 2026.
• Five views on the foundations of data science.. 2026.
• FAIR4prep: FAIR clinical informatics data preprocessing in artificial intelligence applications.. 2026.
• Enabling Next-Generation Mass Spectrometry-Based Proteomics: Standards, Proteoform Resolution, and FAIR, Reproducible, and Quantitative Analysis.. 2026.
• Ten simple rules for achieving computational reproducibility in cognitive neuroscience. 2026.
• Spatio-Temporal COVID-19 Modeling: A Global Systematic Review of Data Integration, Equity, and Lessons for Pandemic Preparedness.. 2026.
• Bioinformatics software development: Principles and future directions. The Innovation Life, 2024.
• A Short Review on Harnessing Bioinformatics for Food Safety: Computational Approaches to Detecting Foodborne Pathogens. Journal of Advanced Health Informatics Research, 2024.
• Open Data to Support CANCER Science-A Bioinformatics Perspective on Glioma Research. Onco, 2021.
• Technical and legal aspects of federated learning in bioinformatics: applications, challenges and opportunities. Frontiers Digit. Health, 2025.
• ‘You should at least ask’. The expectations, hopes and fears of rare disease patients on large-scale data and biomaterial sharing for genomics research. European Journal of Human Genetics, 2016.
• Navigating Copyright and Fair Use in AI Training Data: Legal Challenges and Future Solutions. 3rd International Conference on Advancements in Smart Secure and Intelligent Computing Assic 2025, 2025.
• Copyright protection during the training stage of generative AI: Industry-oriented U.S. law, rights-oriented EU law, and fair remuneration rights for generative AI training under the UN's international governance regime for AI. Computer Law and Security Review, 2024.
• A Fair and Trustworthy Remuneration Framework for AI Model Training Using DLT. 2025 IEEE International Conference on Blockchain and Cryptocurrency Icbc 2025, 2025.

This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.