Co-design for Trustworthy AI: An Interpretable and Explainable Tool for Type 2 Diabetes Prediction Using Genomic Polygenic Risk Scores
New AI tool uses SHAP values to break down polygenic risk scores into specific gene contributions for type 2 diabetes.
A multidisciplinary team from Seoul National University's Graduate School of Data Science has introduced eXplainable PRS (XPRS), a novel tool designed to address a critical limitation in genetic medicine: the lack of interpretability in polygenic risk scores (PRS). PRS are powerful statistical tools that aggregate thousands of genetic variants to quantify an individual's inherited risk for complex diseases like type 2 diabetes, but they traditionally function as a 'black box,' offering a single risk number without explanation. XPRS solves this by applying Shapley Additive Explanations (SHAP), a game theory-based method, to decompose a person's overall PRS into granular, understandable contributions from specific genes and individual single-nucleotide polymorphisms (SNPs). This allows clinicians and patients to see exactly which genetic factors are driving a high-risk prediction, moving from a simple score to actionable biological insight.
Crucially, the researchers didn't just build a technical tool; they embedded it within a rigorous ethical and legal co-design process from the start. They piloted the Council of Europe's Human Rights, Democracy, and the Rule of Law Impact Assessment for AI Systems (HUDERIA) and used the Z-inspection methodology to evaluate XPRS's trustworthiness across legal, medical, ethical, and technical robustness dimensions. This collaborative effort, involving experts in ethics, law, human rights, computer science, and medicine, resulted in a comprehensive 57-page framework of lessons learned. The findings provide a vital blueprint for other researchers and developers aiming to create explainable AI systems for sensitive clinical contexts, ensuring that technological advancement in genomic medicine is matched by responsible and transparent design practices.
- XPRS uses SHAP values to break down a polygenic risk score into specific gene and SNP contributions, making genetic risk interpretable.
- The tool was developed using a co-design 'Z-inspection' process, incorporating ethical, legal, and human rights assessments from the project's inception.
- The resulting 57-page framework provides a multidisciplinary blueprint for building trustworthy, explainable AI in clinical genomics and beyond.
Why It Matters
This bridges the gap between powerful genetic risk prediction and clinical utility by making AI-driven scores transparent and actionable for doctors and patients.