A prospective cohort study in NIH’s All of Us program found that integrating daily step counts and polygenic risk scores (PRS) using machine learning better predicts incident type 2 diabetes (T2D) than either alone.

Why It Matters To Your Practice

• AI-driven models can personalize T2D prevention by combining lifestyle and genetic data.

• Step-count thresholds for reducing risk differ by genetic background, challenging uniform exercise recommendations.

• Clinicians may soon leverage wearables and genomics for tailored risk assessment.

• Understanding these models can improve patient counseling and preventive strategies.

Clinical Implications

• Each additional 1,000 steps/day lowered T2D risk (aHR 0.83), but protective step thresholds were higher for those with increased genetic risk.

• PRS alone strongly predicted T2D (aHR 2.62 per SD increase).

• Combining steps and PRS improved model discrimination (C-index up to 0.867).

• Machine learning methods slightly outperformed traditional Cox models in risk stratification.

Insights

• Physical activity and genetic risk are independent, additive predictors for T2D.

• Penalized Cox and survival SVM models offered the best discrimination, while random survival forests had superior calibration.

• Wearable device data is a scalable resource for precision prevention.

• Results support moving beyond one-size-fits-all activity goals.

The Bottom Line

• Machine learning models combining step counts and genetics can better identify T2D risk, enabling more personalized prevention strategies for patients.