New method improves genetic risk prediction for diabetes complications across diverse populations

For decades, genetic risk prediction has worked best for people of European descent. The polygenic risk scores that doctors use to identify who might develop heart disease, stroke, or kidney failure—particularly among those with type 2 diabetes—were built on data from European populations and simply don't perform as well when applied to South Asian, African, or other ancestry groups. This gap isn't a minor technical problem. It's a fairness issue that could leave millions of patients without reliable guidance about their own health risks.

Researchers at multiple institutions set out to fix this. They developed a new approach called Mondrian Cross-Conformal Prediction, or MCCP, that doesn't try to make genetic scores work equally well across all populations—an impossible task given real biological differences. Instead, it does something different: it tells doctors exactly how confident the prediction is, and flags cases where the model simply cannot make a reliable call. The team tested this method on 19,468 people with type 2 diabetes from the UK Biobank: 17,574 who identified as White British, 1,145 South Asian, and 749 African or Caribbean. They were trying to predict three major complications: stroke, heart attack, and reduced kidney function.

The results were striking in their consistency. When the model was trained on European data from the ADVANCE clinical trial and tested on the three UK Biobank populations, MCCP performed comparably to standard prediction methods across all three groups. For stroke in the White British population, it achieved an AUROC of 0.69—a measure of how well it separated high-risk from low-risk individuals. In the South Asian group, that rose to 0.86. In the African/Caribbean group, it reached 0.87. The same pattern held for heart attack and kidney function predictions. The method wasn't perfect—calibration varied by outcome and ancestry—but it was reliable enough to be clinically useful.

What made MCCP different wasn't raw predictive power. It was transparency. The method allowed researchers to set an acceptable error rate in advance—say, 5 percent—and then identify exactly which patients received confident predictions under that threshold and which ones fell into an "uncertain" or "unpredictable" category. In the White British population at a 5 percent error threshold, MCCP issued confident predictions for 10.3 percent of individuals for stroke, 10.4 percent for heart attack, and 20.7 percent for reduced kidney function. Those numbers were higher in the smaller South Asian and African/Caribbean groups, reflecting the inherent uncertainty of working with less data. But the key insight was that doctors would know which predictions to trust and which ones required additional clinical judgment.

The researchers ran multiple tests to understand where differences came from. They trained models directly in the White British population and tested them in South Asian and African/Caribbean groups, removing the confounding effect of different study designs. They downsampled the larger populations to match the smallest group's size, testing whether sample size alone explained the variation. Across all these scenarios, MCCP maintained comparable performance. The differences between populations reflected real genetic and environmental variation, not methodological bias.

One complication emerged around kidney function prediction. The standard medical formula for estimating kidney function, called CKD-EPI, included a race coefficient—a mathematical adjustment based on the assumption that kidney function differs by race. When researchers removed that coefficient and used the newer race-free version, the number of African/Caribbean patients classified as having reduced kidney function increased from 47 to 62. MCCP's predictions remained valid under both definitions, but the finding highlighted how clinical definitions themselves can embed ancestry-related assumptions that need scrutiny.

The limitations were real. MCCP is computationally more demanding than simple logistic regression, though still manageable for research applications. The method produces predictions it cannot confidently make—and those require explicit clinical protocols that don't yet exist. The South Asian and African/Caribbean cohorts, while substantial, remained much smaller than the White British group, limiting the precision of estimates for rare outcomes. Most critically, the African/Caribbean population in the UK Biobank represents a specific subset of African ancestry with particular admixture patterns; the findings don't automatically transfer to continental African populations or other African diaspora communities.

What the study demonstrated was a path forward. Rather than trying to make one-size-fits-all genetic predictions, the field could embrace uncertainty quantification—explicitly telling clinicians when a prediction is solid and when it isn't. This matters most for the populations that have been historically underserved by genomic research. As the researchers noted, continued validation in larger and more diverse cohorts remains essential, particularly as genomic datasets continue to be dominated by people of European ancestry. The next step is prospective clinical testing: Does this approach actually change how doctors make decisions? Does it improve outcomes? Those questions remain open, but the foundation is now in place.