Integrating Polygenic Risk Improves Generative Forecasting of Disease Trajectories

Area of Science:

• Medical informatics
• Artificial intelligence in healthcare
• Genomics and personalized medicine

Background:

• Predicting lifelong disease sequences is a major medical challenge.
• Existing AI models are limited by cohort size and lack of genetic data integration.
• Generative models offer potential for improved health trajectory prediction.

Purpose of the Study:

• Introduce the Next Health Event (NHE) model, a generative transformer.
• Integrate demographic data, longitudinal BMI, and polygenic risk scores (PRS) with health history.
• Enhance prediction accuracy for future disease diagnoses.

Main Methods:

• Trained a transformer architecture on health trajectories of 7.1 million participants.
• Incorporated demographic data, longitudinal BMI, and PRS for 297 traits.
• Compared NHE model performance against baseline models like XGBoost.

Main Results:

• NHE model achieved higher Top-1 accuracy (25.5%) than XGBoost (22.3%) in predicting the next diagnosis across 129 conditions.
• Polygenic risk scores and longitudinal BMI significantly contributed to predictive power.
• Model demonstrated consistent accuracy for prospective vs. combined outcome reporting (AUROC 0.917).

Conclusions:

• The NHE model establishes a new framework for predictive health by integrating large-scale health histories with genetics.
• Generative models can effectively forecast individual disease pathways.
• PRS and longitudinal BMI are key predictors, while lifestyle information offers limited additional value.