Data Annotation for AI in Life Sciences: Roles, Challenges, and Best Practices

Data annotation in the life sciences is the process of adding structured labels, classifications, and contextual information to raw biological data so that machine learning models can learn from it. This work sits between experimental biology and computational analysis. It requires domain knowledge, careful quality control, and documented decision-making. For students, researchers, analysts, and life-science professionals, understanding annotation roles and processes is essential because model performance depends directly on the quality and consistency of the labels used for training.

This article describes the roles involved in biological data annotation, the common challenges that arise when preparing datasets for machine learning, and practical best practices for building and evaluating annotation workflows. The content is grounded in published evidence from bioinformatics, metabolomics, genomics, and data science literature. A practical checklist for setting up an annotation workflow and evaluating annotation quality is included.

At a Glance

Data annotation for AI in life sciences involves multiple roles, each with distinct responsibilities. The table below summarizes the primary roles, their core tasks, and the key quality considerations for each.

The practical outcome of a well-designed annotation workflow is a dataset that supports reliable model training, reproducible research, and defensible biological conclusions. The remainder of this article details how to achieve that outcome.

The Role of Data Annotation in Biological Machine Learning

Machine learning models in the life sciences depend on labeled training data. Without accurate labels, models cannot learn meaningful patterns. The FAIR Guiding Principles for scientific data management and stewardship emphasize that data should be Findable, Accessible, Interoperable, and Reusable, with specific attention to enabling machines to automatically find and use data in addition to supporting reuse by individuals. Annotation is the process that makes raw biological data machine-actionable in this way.

In practice, annotation takes many forms depending on the data type. Genomic sequences require variant classification, gene structure annotation, and functional labeling. Mass spectrometry data requires peak identification and metabolite annotation. Single-cell transcriptomic data requires cell type labeling. Clinical data requires diagnosis coding and outcome labeling. Each of these tasks demands different expertise and different quality control approaches.

The scale of annotation work in modern life sciences is substantial. For example, an integrated transcriptomic cell atlas of human neural organoids combined 36 single-cell datasets spanning 26 protocols to create a resource of more than 1.7 million cells. Building such an atlas required consistent cell type annotation across heterogeneous datasets generated by different laboratories using different protocols. The effort demonstrates that annotation is not a peripheral task but a central component of modern biological data science.

Similarly, a curated multimodal single-cell atlas of the human thymus required spatially resolved analysis, integration of spatial transcriptomics, and high-resolution multiplex imaging data. The authors developed a toolkit for cross-platform imaging data analysis and annotation that can be applied to any tissue. This example shows that annotation tools and frameworks are becoming reusable infrastructure instead of one-off efforts.

Core Principles of Biological Data Annotation

Several principles guide effective annotation in the life sciences. These principles apply across data types and research contexts.

Annotation Requires Domain Expertise

Biological data annotation is not a purely mechanical task. It requires understanding the biology underlying the data. A survey on deep learning tools dealing with data scarcity notes that manual labeling typically involves human annotators with a vast background of knowledge, and that this annotation process is costly, time-consuming, and error-prone. The cost and error rate are directly related to the complexity of the biological questions being asked.

For example, annotating TP53 variations in human cancers requires understanding the functional significance of specific mutations. The IARC TP53 Database compiles occurrence and phenotype data on TP53 germline and somatic variations linked to human cancer. Analysis of this database showed that genomic studies revealed mutations occurring outside the DNA-binding domain in specific cancer types, findings that were overlooked in earlier data. This example illustrates that annotation quality depends on both the data available and the biological knowledge applied to interpret it.

Annotation Must Be Documented and Reproducible

Annotation decisions must be recorded so that others can understand how labels were assigned. This includes documenting the criteria used, the version of reference databases consulted, and the identity of the annotators. Enabling annotation provenance in bioinformatics workflow applications is an established concern in the field. Provenance tracking allows errors to be traced to their source and corrected systematically instead of through ad hoc fixes.

The FAIR principles provide a framework for this documentation. The principles emphasize that data should be findable, accessible, interoperable, and reusable, with specific attention to machine-actionability. Annotation adds value to data only when the annotation process itself is transparent and reproducible.

Annotation Quality Must Be Measured

Quality cannot be assumed. It must be measured through structured processes. Inter-annotator agreement, agreement with reference standards, and consistency checks are all measurable. A comprehensive review of federated learning challenges from a data preparation viewpoint notes that high-quality data is characterized by completeness, consistency, accuracy, representativeness, and homogeneity. These characteristics apply equally to annotated datasets used for training machine learning models.

In the context of federated learning, heterogeneity caused by inconsistent labeling, missing values, and class imbalances across data nodes degrades model performance and may lead to overfitting, underfitting, and erroneous decision-making. The same issues arise in centralized biological datasets when annotation standards are not applied consistently.

Practical Workflow for Setting Up Data Annotation

Setting up a data annotation workflow requires planning across several dimensions. The following steps provide a structured approach.

Step 1: Define the Annotation Task and Schema

Before any labeling begins, the annotation task must be precisely defined. What biological question is being addressed? What labels are needed? What is the granularity of the classification? For example, cell type annotation in single-cell data requires a defined ontology of cell types. Variant annotation requires a defined schema for variant classes and functional impact.

The annotation schema should be documented and versioned. Changes to the schema should be tracked, and the impact of schema changes on existing labels should be assessed.

Step 2: Select and Prepare the Data

Raw data must be organized and cleaned before annotation. This includes removing technical artifacts, standardizing file formats, and recording metadata. The NCBI Data Resources provide access to a wide range of biological databases that can serve as reference standards for annotation. EMBL-EBI Training offers educational resources on data management and bioinformatics that can support this preparation phase.

Data preparation also includes considering data scarcity. A survey on deep learning tools dealing with data scarcity notes that many applications have small or inadequate data to train deep learning frameworks, and that manual labeling is often needed to provide labeled data. When data is scarce, annotation workflows must prioritize the most informative samples.

Step 3: Recruit and Train Annotators

Annotators must have appropriate domain expertise for the task. For complex biological annotation, this typically means graduate-level training in the relevant field. For simpler tasks, trained technicians may be sufficient, but they must work under close supervision from domain experts.

Training should include the annotation schema, examples of correct and incorrect labels, and practice sets with feedback. The training process should be documented so that new annotators can be onboarded consistently.

Step 4: Establish Quality Control Procedures

Quality control should be built into the workflow from the start, not added after annotation is complete. This includes:

• Defining a gold standard set of samples with known correct labels
• Measuring inter-annotator agreement on a subset of samples
• Establishing thresholds for acceptable agreement
• Defining procedures for resolving disagreements
• Sampling annotated data for review by senior experts

The thresholds and procedures should be documented and applied consistently.

Step 5: Implement the Annotation Process

The annotation process itself should be managed through a system that tracks progress, records decisions, and supports review. This can be a commercial annotation platform, an open-source tool, or a custom-built system. The key requirement is that the system must support provenance tracking and audit trails.

Step 6: Evaluate Annotation Quality

After annotation is complete, quality must be evaluated systematically. This includes measuring agreement with the gold standard, checking for systematic biases, and assessing coverage of the annotation schema. The evaluation results should be documented and used to improve future annotation efforts.

Step 7: Document and Release the Annotated Dataset

The annotated dataset should be documented with clear descriptions of the annotation schema, the quality control procedures, and the limitations of the labels. The FAIR principles provide guidance on making data findable, accessible, interoperable, and reusable. The NIH Genomic Data Sharing Policy provides additional requirements for genomic data generated with NIH funding.

Options and Tradeoffs in Annotation Approaches

Different annotation approaches have different strengths and weaknesses. The choice of approach depends on the data type, the available expertise, the scale of the task, and the quality requirements.

Manual Annotation

Manual annotation by domain experts produces the highest quality labels for complex biological tasks. However, it is slow, expensive, and difficult to scale. A survey on deep learning tools dealing with data scarcity notes that manual labeling is costly, time-consuming, and error-prone, even when performed by knowledgeable annotators.

Manual annotation is appropriate when the annotation task requires nuanced biological judgment, when the data volume is manageable, and when the cost of errors is high. For example, functional annotation of phosphorylation sites requires integrating proteomic, structural, regulatory, and evolutionary evidence. A study that manually curated 112 datasets of phospho-enriched proteins and re-analyzed 6,801 proteomics experiments created a reference phosphoproteome containing 119,809 human phosphosites. The manual curation effort was substantial but produced a high-value resource.

Semi-Automated Annotation

Semi-automated annotation combines automated pre-annotation with human review. Automated tools can propose labels based on pattern matching, database lookup, or preliminary models. Human annotators then review and correct the proposed labels.

This approach can significantly increase throughput while maintaining quality. However, it introduces the risk that annotators will accept incorrect automated labels without sufficient scrutiny. Validation of automated tools against human-labeled gold standards is essential before deployment.

Fully Automated Annotation

Fully automated annotation uses algorithms to assign labels without human review. This approach is appropriate when the annotation task is well-defined, the reference databases are comprehensive, and the cost of errors is low.

However, fully automated annotation has significant limitations. A study on labeling network traffic found that all current methods for labeling suffer from fundamental drawbacks regarding the quality, volume, and speed of the resulting dataset. The same issues apply to biological data. Automated annotation methods may miss rare or novel patterns that require human expertise to identify.

Crowdsourced Annotation

Crowdsourced annotation distributes labeling tasks to a large number of non-expert workers. This approach can be cost-effective for simple tasks but is problematic for complex biological annotation. The study on labeling network traffic noted that methods incorporating non-experts in the labeling process of real traffic with the help of visual and statistical tools still suffer from fundamental drawbacks. For biological data, the domain knowledge required for accurate annotation makes crowdsourcing impractical for most tasks.

Community Curation

Community curation involves the collective effort of researchers in a field to annotate and maintain shared resources. The Global Natural Products Social Molecular Networking platform provides an example of this approach. The platform is an open-access knowledge base for community-wide organization and sharing of raw, processed, or identified tandem mass spectrometry data. Crowdsourced curation of freely available community-wide reference mass spectrometry libraries underpins improved annotations.

Community curation can be effective when the community has strong incentives to contribute and when clear standards are established. However, it requires ongoing governance to maintain quality and consistency.

Observations and Measurements in Annotation Quality

Measuring annotation quality requires structured observation and documentation. The following measurements are commonly used.

Inter-Annotator Agreement

Inter-annotator agreement measures the extent to which different annotators assign the same labels to the same data. Low agreement indicates that the annotation schema is ambiguous or that annotators lack sufficient training. High agreement indicates consistency but does not guarantee accuracy, since all annotators may share the same misconceptions.

Agreement with Gold Standard

Agreement with a gold standard measures the accuracy of annotations against a reference set of samples with known correct labels. The gold standard should be established by senior experts using the best available evidence. Disagreement with the gold standard indicates annotation errors that need correction.

Label Distribution Analysis

The distribution of labels across categories should be examined for unexpected patterns. For example, if one cell type label is assigned much more frequently than expected based on biological knowledge, this may indicate a systematic bias in the annotation process.

Error Pattern Analysis

When errors are identified, they should be analyzed for patterns. Are errors concentrated in specific data types? Are they associated with specific annotators? Are they related to specific parts of the annotation schema? Understanding error patterns enables targeted improvements to the annotation process.

Model Performance Correlation

The ultimate test of annotation quality is the performance of models trained on the annotated data. If model performance is poor on specific classes or data types, this may indicate annotation problems. Close feedback loops between model development and annotation quality are essential.

Records and Documentation Requirements

Documentation is a critical component of annotation workflows. The following records should be maintained.

Annotation Schema Documentation

The annotation schema should be documented with clear definitions of each label, inclusion and exclusion criteria, and examples. Schema versions should be tracked, and changes should be documented with rationales.

Annotator Training Records

Records of annotator training should include the training materials used, the dates of training, and the results of any training assessments. These records support quality assurance and enable consistent onboarding of new annotators.

Annotation Decision Logs

Each annotation decision should be logged with the data sample identifier, the assigned label, the annotator identity, the date, and any notes about the decision. This log supports error tracing and dispute resolution.

Quality Control Records

Quality control records should document the samples reviewed, the review results, the agreement measurements, and any corrective actions taken. These records provide evidence of annotation quality for downstream users.

Dataset Version Records

Annotated datasets should be versioned, with records of what changed between versions and why. Version records support reproducibility and enable users to understand the history of the data.

Common Failure Patterns in Annotation Projects

Understanding common failure patterns helps annotation projects avoid predictable problems.

Schema Ambiguity

When the annotation schema is ambiguous, annotators interpret labels differently, leading to inconsistent annotations. This problem is often detected through low inter-annotator agreement. The solution is to refine the schema with clearer definitions and more examples.

Annotator Drift

Annotator drift occurs when annotators become less consistent over time as they gain experience or become fatigued. Regular quality checks and refresher training can mitigate this problem.

Automation Bias

When automated pre-annotation is used, annotators may accept incorrect automated labels without sufficient scrutiny. This problem is particularly dangerous because it can systematically propagate errors. Validation of automated tools and periodic audits of accepted labels are essential.

Reference Database Obsolescence

Biological reference databases change over time as new knowledge is accumulated. Annotations based on outdated references may be incorrect. The IARC TP53 Database analysis demonstrated that genomic studies revealed findings overlooked in earlier data, highlighting the importance of updating annotations as knowledge advances.

Class Imbalance

When some classes are rare in the training data, models may perform poorly on those classes. The survey on deep learning tools dealing with data scarcity notes that imbalanced datasets are a major challenge. Annotation workflows should monitor class distributions and consider oversampling or other strategies for rare classes.

Provenance Loss

When provenance information is not recorded, errors cannot be traced to their source, and the dataset cannot be updated systematically. Provenance loss undermines the long-term value of annotated datasets.

Limitations of Data Annotation in Life Sciences

Data annotation has inherent limitations that must be acknowledged.

Annotation Is Interpretive

Biological annotation involves interpretation, beyond measurement. Different experts may legitimately disagree about the correct label for a given data sample. The annotation process should acknowledge this interpretive dimension and document the basis for decisions.

Reference Standards Are Incomplete

Biological knowledge is incomplete. Reference databases may not contain all relevant information, and some biological entities may not yet be characterized. The NetID approach to metabolomics annotation illustrates this limitation. The study found that most liquid chromatography-high-resolution mass spectrometry peaks remain unidentified, and the global network optimization approach was developed to improve annotation coverage. Even with advanced methods, some peaks remain unannotated.

Annotation Quality Is Context-Dependent

The appropriate level of annotation detail depends on the intended use of the data. A dataset annotated for one purpose may not be suitable for another. Annotation workflows should document the intended use cases and the limitations of the annotations for other purposes.

Annotation Is Never Complete

Biological knowledge evolves, and annotations must be updated as new information becomes available. The concept of living data, introduced in the context of the Global Natural Products Social Molecular Networking platform, involves continuous reanalysis of deposited data. This approach acknowledges that annotation is an ongoing process instead of a one-time effort.

Safety and Regulatory Context

Data annotation in the life sciences operates within a regulatory and ethical framework that varies by data type and jurisdiction.

Genomic Data Sharing

The NIH Genomic Data Sharing Policy establishes expectations for the sharing of genomic data generated with NIH funding. The policy addresses data submission, data access, and data security. Annotation workflows for genomic data must comply with these requirements, including appropriate handling of potentially identifiable data.

Data Privacy and Security

Biological data may contain sensitive information, including genetic data that could identify individuals. Annotation workflows must include appropriate privacy and security controls. This includes access controls, data encryption, and procedures for handling data breaches.

Ethical Considerations

Annotation decisions can have ethical implications. For example, variant classification can affect clinical decisions. The study on AI and the transformation of clinical reasoning highlights the importance of understanding how AI tools, including those trained on annotated data, affect clinical practice. Annotation workflows should consider the potential consequences of labeling decisions.

Professional Escalation Criteria

When annotation decisions have significant consequences, escalation procedures should be in place. For example, if an annotator encounters a data sample that does not fit the annotation schema, or if there is disagreement about a label that cannot be resolved through normal procedures, the issue should be escalated to senior experts. Escalation criteria should be defined in advance and documented in the annotation protocol.

Best Practices for Annotation Quality Assurance

The following best practices are drawn from the evidence reviewed in this article.

Establish Clear Standards

Clear annotation standards are the foundation of quality. The FAIR principles provide a framework for making data findable, accessible, interoperable, and reusable. The NIH Genomic Data Sharing Policy provides specific requirements for genomic data. Annotation workflows should adopt and document relevant standards.

Use Multiple Annotators

Using multiple annotators for the same data enables measurement of inter-annotator agreement and identification of ambiguous cases. The study on the functional landscape of the human phosphoproteome involved manual curation of 112 datasets, demonstrating the value of careful human curation even in large-scale projects.

Validate Automated Tools

Automated annotation tools must be validated against human-labeled gold standards before deployment. The study on mechanisms for automatic training data labeling for machine learning addresses approaches for automating labeling while maintaining quality. Validation should be ongoing, not a one-time event.

Monitor Quality Continuously

Quality monitoring should be continuous, not episodic. Regular sampling of annotated data for review, ongoing measurement of inter-annotator agreement, and analysis of error patterns all support continuous quality improvement.

Document Everything

Documentation supports reproducibility, error tracing, and continuous improvement. The study on enabling annotation provenance in bioinformatics workflow applications emphasizes the importance of provenance for scientific workflows.

Build Feedback Loops

Feedback loops between annotators, quality assurance reviewers, and model developers improve both annotation quality and model performance. When models perform poorly on specific classes, this information should flow back to annotators to identify potential annotation issues.

Professional Escalation Criteria

Annotation projects should define clear escalation criteria for situations that require senior expert involvement. The following situations warrant escalation.

Schema Gaps

When data samples do not fit the annotation schema, the schema may need revision. Escalation to senior experts is appropriate to determine whether the schema should be expanded or whether the samples should be excluded.

Persistent Disagreement

When annotators consistently disagree on specific types of samples, escalation is appropriate to resolve the underlying ambiguity. Persistent disagreement may indicate that the annotation schema needs refinement or that additional training is needed.

Potential Data Quality Issues

When annotation reveals potential problems with the underlying data, such as technical artifacts or sample mislabeling, escalation is appropriate to determine whether the data should be corrected or excluded.

Ethical or Regulatory Concerns

When annotation decisions raise ethical or regulatory concerns, such as potential identification of research participants or conflicts with data sharing policies, escalation to appropriate oversight bodies is required.

Model Performance Problems

When models trained on annotated data perform poorly in ways that suggest annotation problems, escalation is appropriate to investigate and correct the underlying issues.

Frequently Asked Questions

What qualifications do data annotators need in the life sciences?

Data annotators in the life sciences need domain expertise relevant to the data type they are labeling. For complex tasks such as variant classification or cell type annotation, graduate-level training in the relevant field is typically required. For simpler tasks, trained technicians may be sufficient, but they must work under close supervision from domain experts. The survey on deep learning tools dealing with data scarcity notes that manual labeling typically involves human annotators with a vast background of knowledge, reflecting the expertise required for high-quality annotation.

How is annotation quality measured?

Annotation quality is measured through several complementary approaches. Inter-annotator agreement measures consistency between annotators. Agreement with a gold standard measures accuracy against a reference set of samples with known correct labels. Label distribution analysis examines whether labels are distributed as expected based on biological knowledge. Error pattern analysis identifies systematic problems in the annotation process. Model performance correlation assesses whether models trained on annotated data perform as expected.

What is the difference between data annotation and data curation?

Data annotation adds labels and classifications to raw data, while data curation involves organizing, standardizing, and maintaining data resources. The two activities are closely related and often overlap. Curation includes tasks such as format standardization, metadata management, and version control. Annotation adds the domain-specific labels that make data useful for machine learning. Both activities are essential for building high-quality training datasets.

How do automated annotation tools compare to human annotators?

Automated annotation tools can process data much faster than human annotators and can be more consistent in applying defined rules. However, they may miss rare or novel patterns that require human expertise. The study on labeling network traffic found that all current methods for labeling suffer from fundamental drawbacks regarding the quality, volume, and speed of the resulting dataset. For biological data, automated tools are most effective when combined with human review in a semi-automated workflow.

What is the role of reference databases in annotation?

Reference databases provide the standards against which annotations are made. For example, the NCBI Data Resources provide access to a wide range of biological databases that can serve as reference standards. The IARC TP53 Database compiles occurrence and phenotype data on TP53 germline and somatic variations linked to human cancer. Reference databases must be kept current, as annotations based on outdated references may be incorrect.

How should annotation disagreements be resolved?

Annotation disagreements should be resolved through structured procedures. The first step is to document the disagreement and the basis for each annotator's decision. The second step is to review the annotation schema to determine whether it is ambiguous. The third step is to escalate unresolved disagreements to senior experts. The resolution should be documented and used to improve the annotation schema and training materials.

What are the main challenges in annotating metabolomics data?

Metabolomics annotation faces the challenge that most peaks in liquid chromatography-high-resolution mass spectrometry data remain unidentified. The NetID study found that most peaks remain unidentified and developed a global network optimization approach to improve annotation coverage and accuracy. The Global Natural Products Social Molecular Networking platform addresses this challenge through community curation of reference mass spectrometry libraries. Metabolomics annotation requires integrating measured masses, retention times, and tandem mass spectrometry fragmentation patterns.

How does data annotation affect model performance?

Data annotation quality directly affects model performance. The comprehensive review of federated learning challenges notes that machine learning model accuracy, generalization, and reliability are greatly affected by training data quality. High-quality data characterized by completeness, consistency, accuracy, representativeness, and homogeneity enables meaningful pattern learning and robust prediction. Inconsistent labeling, missing values, and class imbalances degrade model performance and may lead to overfitting, underfitting, and erroneous decision-making.

Related Bioinformatics Guides

• The UK Biobank: Managing Massive Biological Datasets
• Machine Learning for Predicting T-Cell Epitope Immunogenicity
• Machine Learning Classification of Antimicrobial Resistance Genes
• Machine Learning for Variant Effect Prediction on Protein Stability
• Machine Learning for Predicting Host-Virus Protein-Protein Interaction Networks

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.
• Metabolite discovery through global annotation of untargeted metabolomics data.. Nature methods, 2021.
• The FAIR Guiding Principles for scientific data management and stewardship.. Scientific data, 2016.
• TP53 Variations in Human Cancers: New Lessons from the IARC TP53 Database and Genomics Data.. Human mutation, 2016.
• The functional landscape of the human phosphoproteome.. Nature biotechnology, 2020.
• An integrated transcriptomic cell atlas of human neural organoids.. Nature, 2024.
• Sharing and community curation of mass spectrometry data with Global Natural Products Social Molecular Networking.. Nature biotechnology, 2016.
• A spatial human thymus cell atlas mapped to a continuous tissue axis.. Nature, 2024.
• Updated unified phylogenetic classification system and revised nomenclature for Newcastle disease virus.. Infection, genetics and evolution : journal of molecular epidemiology and evolutionary genetics in infectious diseases, 2019.
• AI and the transformation of clinical reasoning: A classroom perspective.. 2026.
• [Construction of an artificial intelligence-enabled closed-loop teaching model for Microbiology Experiments].. 2026.
• Fung-AI: An AI/ML-driven pipeline for antifungal peptide discovery.. 2026.
• Will artificial intelligence change medical life? Bridging the gap between innovation and medical student adoption: a cross-sectional study among university medical students.. 2026.
• Development and preliminary evaluation of a German version of the active communication education group aural rehabilitation training program.. 2026.
• The impact of an artificial intelligence (AI)-assisted education program on mental health, social support and quality of life in older individuals with head and neck cancer: study protocol of a randomized controlled trial.. 2026.
• Data Labeling: An Empirical Investigation into Industrial Challenges and Mitigation Strategies. International Conference on Product Focused Software Process Improvement, 2020.
• Mechanisms for Automatic Training Data Labeling for Machine Learning. International Conference on Interaction Sciences, 2019.
• Comprehensive review of federated learning challenges: a data preparation viewpoint. Journal of Big Data, 2025.
• A survey on deep learning tools dealing with data scarcity: definitions, challenges, solutions, tips, and applications. Journal of Big Data, 2023.
• Datasets are not Enough: Challenges in Labeling Network Traffic. Computers & security, 2021.
• Enabling annotation provenance in bioinformatics workflow applications. Lecture Notes in Computer Science Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics, 2010.
• Bioinformatics for spermatogenesis: Annotation of male reproduction based on proteomics. Asian Journal of Andrology, 2013.
• Bioinformatics issues for automating the annotation of genomic sequences.. Genome Informatics International Conference on Genome Informatics, 2001.

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