The Hidden Cost of Bias in Training Data for Polygenic Risk Scores

Polygenic risk scores (PRS) fail for most global populations because they are trained predominantly on genetic data from individuals of European ancestry. This creates a precision medicine paradox where 'personalized' predictions are only accurate for a narrow demographic slice.

Training data bias is systemic. Major biobanks like UK Biobank and genome-wide association studies (GWAS) have historically over-sampled European cohorts. This means the single nucleotide polymorphisms (SNPs) identified as significant for disease risk are not generalizable across ancestries, leading to poor predictive performance in African, Asian, and Indigenous populations.

The technical consequence is model collapse. A PRS model trained on non-diverse data exhibits high variance in effect size estimates when applied to new populations. For example, a score predicting coronary artery disease risk developed in Europeans can be statistically useless for individuals of South Asian descent, as highlighted in studies from the Broad Institute.

This bias perpetuates health disparities. Deploying these scores in clinical decision support systems or for patient stratification in trials risks misdiagnosis and unequal allocation of preventive care. It transforms a tool for equity into an engine of exclusion, a core challenge in building ethical AI for genomics.

Biased training data directly sabotages the clinical utility of polygenic risk scores (PRS) by creating models that fail for anyone not represented in the original cohort. PRS models, built using statistical frameworks like Plink or REGENIE, calculate disease risk by summing the effects of thousands of genetic variants; if those variant-disease associations are derived almost exclusively from European-ancestry genomes, the resulting scores are statistically invalid for other groups.

The problem is systemic, not incidental. Landmark biobanks like the UK Biobank are overwhelmingly composed of participants of European descent. This creates a foundational skew in the genome-wide association studies (GWAS) that power PRS, meaning the 'risk' alleles identified are specific to one population's genetic architecture and linkage disequilibrium patterns.

The consequence is predictive failure. A PRS for type 2 diabetes trained on European data can misclassify risk for individuals of South Asian or African ancestry by over 50%. This isn't a minor error—it's a clinical liability that perpetuates health disparities by denying accurate preventative care to underrepresented groups.

Evidence: A 2023 study in Nature demonstrated that PRS accuracy, measured by AUC, drops by an average of 0.15 to 0.30 when applied to populations not represented in the training data. This performance collapse renders the scores clinically useless and ethically dangerous.

Data availability is not a justification for bias; it is a solvable engineering problem. Models like polygenic risk scores (PRS) trained on non-diverse cohorts, such as the UK Biobank, fail for underrepresented populations because they learn spurious correlations specific to the training data's genetic ancestry. This creates a technical debt of inequity that is expensive to fix post-deployment.

The fallacy confuses quantity with representativeness. A model trained on a million genomes from a single population has less generalizable power than one trained on 100,000 genomes from five ancestries. Federated learning frameworks like NVIDIA FLARE or OpenFL enable training across global biobanks without centralizing data, directly addressing the privacy and diversity challenge.

Bias is a feature, not a bug, of the training pipeline. Standard practices in genomic AI, like using principal component analysis (PCA) for population stratification, often bake in ancestry as a confounding variable. Techniques like counterfactual fairness and adversarial debiasing must be integrated during model development, not as an afterthought.

Evidence: A 2023 study in Nature showed a PRS for coronary artery disease trained on European data had an AUC drop of over 0.15 when applied to South Asian populations. This performance gap represents a direct clinical risk for millions of patients. Solving this requires a commitment to deliberate data acquisition and techniques like synthetic data generation to augment underrepresented groups, a core component of our work in synthetic data for clinical trials.

Polygenic risk scores (PRS) trained on non-diverse genomic data produce invalid clinical predictions for underrepresented populations, triggering imminent regulatory action. The foundational models for these scores rely on datasets like the UK Biobank, which are overwhelmingly of European ancestry. This creates a genomic data gap that systematically misrepresents the genetic architecture of other populations, rendering the scores clinically useless and ethically indefensible.

The core technical failure is population stratification, not algorithmic bias. Standard PRS models assume genetic effects are consistent across ancestries, a flawed premise that amplifies health disparities. A score predicting a 10% risk of Type 2 Diabetes for a European-ancestry individual could indicate a 2% or 25% risk for someone of African or Asian descent, making the tool dangerous for clinical decision-making.

Regulatory bodies like the FDA and EMA will mandate diversity in training data as a precondition for approval. The EU AI Act already classifies high-risk AI systems in healthcare, requiring rigorous bias and fairness auditing. Developers using platforms like TensorFlow Extended (TFX) or MLflow for MLOps must now integrate ancestry-aware validation suites and demonstrate generalizability across global populations, a core tenet of responsible AI TRiSM.