Dataset Lineage in Genomic and Clinical AI Pipelines
Holding
Genomic and clinical AI cannot be governed by final dataset names. It must be governed by lineage: consent, accession, extraction, processing, variant calling, phenotype linkage, feature generation, and model use.
Authority
NIH genomic data policy expects responsible sharing and use of genomic data, including controlled-access conditions. HIPAA requires safeguards for identifiable health information. These duties become technical when data are copied, transformed, joined, or exported.
Issue
The failure mode is lawful access followed by unlawful or unsupported reuse. A dataset may be permitted for one research purpose, but later features, embeddings, or derived cohorts may travel into a different project without the original use restriction attached.
Resolution
Lineage should travel with derived artifacts. The platform should attach dataset identifiers, consent terms, repository terms, project authorization, preprocessing version, and export restrictions to every materialized table, vector store, and model training run.
Evidence Package
The record should show accession identifiers, data use terms, approved users, transformation logs, derived artifact inventory, training run metadata, and export review. If a model was trained on genomic data, the system should be able to identify which rights and restrictions followed it.