Ethics, Bias, and Responsible Use in Genetics

Written By Kamayani Gupta

Ancestry Bias and Underrepresentation in Training Data

AI models in genetics often suffer from ancestry bias due to disproportionate representation of European populations. In polygenic risk prediction, deep learning frameworks that disentangle ancestry from phenotype-relevant signals have shown improved performance across diverse and admixed groups, such as in Alzheimer’s disease prediction models.

Moreover, the tool PhyloFrame uses population genomics and gene-network data to mitigate ancestral bias in predictive models. It improves disease risk prediction across multiple cancer types [breast, thyroid, uterine] in diverse populations by accounting for enhanced allele frequencies in underrepresented groups.

Recognizing these biases is essential. A recent scoping review found that 67% of clinical ML studies failed to meaningfully address racial bias. Genetic risk scores and AI predictions are probabilistic summaries, not definitive diagnoses. Misinterpretation, especially in complex traits, can lead to unnecessary anxiety, over-treatment, or false reassurance. A 2024 UK report on medical device equity notes concerns over polygenic risk scores exacerbating healthcare disparities because of ancestry bias.

Moreover, AI algorithms that infer sensitive information, such as race or genetic ancestry, can produce divergent outcomes, sabotaging fairness and exacerbating inequity in health recommendations.

Data Privacy and Reidentification Risks

Genetic data are inherently identifiable. Technologies combining genetic and public data sets can reidentify supposedly anonymized individuals. This raises serious privacy concerns. Federal policies like the Common Rule, GINA, and HIPAA offer protections, but gaps remain, especially for life insurance, military personnel, and family members of tested individuals.

Geneticists must apply strong governance, secure consent, and data minimization to prevent misuse and maintain public trust.

Regulatory Landscape: FDA, CLIA, and AI-Based Genetic Tests

The FDA now classifies high-risk laboratory-developed tests (LDTs), including many genomic assays, as medical devices subject to premarket review and validation, aiming to close regulatory gaps. This shift has prompted pushback from stakeholder groups like the Association for Molecular Pathology, which challenges FDA overreach on LDT oversight.

FDA is also developing a risk-based, lifecycle regulatory framework for AI/ML-based medical devices, including genetic software, aligning with IMDRF principles. Finally, CLIA continues to regulate laboratory testing quality; genetic AI tools used in labs must meet both CLIA and FDA standards where applicable.

Best Practices for Validation and Reporting

To promote trustworthy AI in genetics, adhere to the following best practices:

  • Diverse validation across populations, ancestries, and technologies. Use external cohorts to detect performance drift.

  • Transparent reporting of model intent, training data demographics, and known limitations.

  • Fairness assessments stratified by demographic subgroups.

  • Uncertainty estimation and clear communication about predictive confidence.

  • Stakeholder engagement including ethicists and community representatives to guide responsible use.

These practices are in line with contemporary recommendations for responsible AI deployment in medicine.

Resources

Kamayani Gupta
Previous

Clinical Applications and Explainability