Cardiometabolic risk estimation using exposome data and machine learning.

Area of Science:

• Environmental Health
• Computational Biology
• Preventive Medicine

Background:

• The human exposome, encompassing all lifetime exposures, significantly impacts health outcomes alongside genetics.
• Exposome factors are increasingly recognized as critical contributors to major diseases like cardiovascular disease (CVD) and type 2 diabetes (T2D).
• Personalized risk assessment using exposome data offers a promising avenue for early disease detection and prevention.

Purpose of the Study:

• To develop and evaluate a novel, fair machine learning (ML) model for predicting CVD and T2D risk.
• The model utilizes readily available exposome factors and is validated across multi-center cohorts.
• Fairness criteria ensured consistent model performance across diverse demographic subgroups.

Main Methods:

• Utilized UK Biobank data with 5,348 CVD and 1,534 T2D cases, matched with controls.
• Incorporated 109 exposome variables (physical, environmental, lifestyle, mental health, sociodemographic, early-life).
• Employed XGBoost ML model, comparing its performance against an integrative ML model and the Framingham risk score for CVD; assessed for bias and interpreted using SHAP.

Main Results:

• The exposome-based ML model achieved comparable performance to the integrative ML model (ROC-AUC 0.78±0.01 for CVD, 0.77±0.01 for T2D).
• For CVD risk, the exposome model outperformed the traditional Framingham risk score.
• The model demonstrated no significant bias across sex, ethnicity, or age subgroups.

Conclusions:

• Key exposome factors for CVD and T2D risk include daytime naps, education level, smoking history, fatigue, and work status.
• This study highlights the potential of exposome-driven machine learning for fair and effective CVD and T2D risk assessment.
• Exposome-based ML provides a valuable tool for personalized, early risk identification and disease prevention strategies.