Multimodal Data Integration Improves Disease Risk Prediction in the UK Biobank

Area of Science:

• Computational biology
• Genomics
• Medical informatics

Background:

• Family health history is crucial for assessing chronic disease risk.
• Integrating electronic health records (EHRs) with genetic data can enhance disease risk prediction by capturing diverse patient factors.
• Current models often struggle to comprehensively integrate multimodal patient data.

Purpose of the Study:

• To develop and evaluate ALIGATEHR-Gen, a novel graph attention network for improved disease risk prediction.
• To integrate multimodal patient data, including genetic information, EHRs, and medical ontologies.
• To enhance patient representation by incorporating genetically inferred relationships and disease ontology embeddings.

Main Methods:

• ALIGATEHR-Gen utilizes a graph attention network architecture.
• The model integrates multimodal patient data: genetic information, diagnosis codes, and demographics.
• External medical ontology knowledge and genetically inferred first-degree relationships are incorporated.
• Model performance is evaluated on 118 diseases using UK Biobank data.

Main Results:

• ALIGATEHR-Gen outperforms state-of-the-art baseline models by an average of at least 6% in disease risk prediction.
• The model demonstrates effective distinction of patient subgroups for fibrotic and related diseases based on clinical and genetic features.
• Unified patient representations enhance predictive accuracy.

Conclusions:

• ALIGATEHR-Gen shows significant potential for advancing predictive and interpretable modeling in healthcare.
• The integration of multimodal data, including genetic and EHR information, is key to improving disease risk prediction.
• This approach offers a pathway to more personalized and accurate patient risk stratification.