An informed machine learning based environmental risk score for hypertension in European adults

Area of Science:

• Environmental Health
• Computational Biology
• Epidemiology

Background:

• The exposome framework aims to understand cumulative environmental exposures' health impacts.
• Existing methods face challenges like multicollinearity, non-linearity, and confounding.
• SEANN (Summary Effect Adjusted Neural Network) is introduced to address these limitations.

Purpose of the Study:

• To develop a novel approach integrating domain knowledge with neural networks for analyzing hypertension risk factors.
• To improve the analysis and interpretation of environmental exposures' effects on hypertension.
• To compare a SEANN-informed model with an agnostic deep neural network model.

Main Methods:

• Utilized data from 18,337 adults (40-65y) in the GCAT cohort, analyzing 53 environmental factors.
• Computed two deep neural network-based environmental risk scores for hypertension prevalence: one informed by SEANN and pooled effect sizes, and an agnostic counterpart.
• Employed Shapley values to extract and compare exposure-outcome relationships learned by both models.

Main Results:

• Both agnostic NN and SEANN models achieved similar predictive performance (AUC 0.7).
• SEANN demonstrated substantial improvements in the scientific validity of learned relationships, aligning better with existing literature.
• Variables directly informed by SEANN were closer to literature findings, and non-informed variables showed adjusted associations more consistent with previous studies; mean delta SHAP distance was 6 times lower with SEANN.

Conclusions:

• SEANN offers added value over conventional, purely data-driven machine learning approaches in environmental health research.
• By incorporating literature-based effect sizes, SEANN enhances the disentanglement of complex exposure effects on hypertension.
• The study highlights SEANN's potential for more accurate and interpretable analysis of exposome data.