Gene expression knowledge graph for patient representation and diabetes prediction.

Area of Science:

• Biomedical Informatics
• Machine Learning
• Genomics

Background:

• Diabetes mellitus is a significant global health concern impacting millions.
• Gene expression data analysis shows potential for improved diabetes prediction.
• Integrating diverse gene expression datasets is challenging due to limited patient numbers and data incompatibility.

Purpose of the Study:

• To develop a novel method for integrating multiple, incompatible gene expression datasets for enhanced diabetes prediction.
• To leverage knowledge graphs for biomedical data integration and learning uniform patient representations.
• To improve the accuracy of diabetes prediction by combining gene expression data with domain-specific knowledge.

Main Methods:

• Utilized knowledge graphs for integrating multiple gene expression datasets and domain-specific knowledge (protein functions, interactions).
• Explored various strategies and knowledge graph (KG) embedding methods to generate uniform patient vector representations.
• Employed these representations as input for a machine learning classifier to predict diabetes.

Main Results:

• Demonstrated significant improvements in diabetes prediction accuracy through the integration of multiple datasets and knowledge graphs.
• Achieved weighted F1-score improvements of up to 13% in diabetes prediction.
• Validated the efficacy of the proposed approach in handling incompatible gene expression datasets.

Conclusions:

• The proposed knowledge graph-based integration method effectively addresses challenges in combining heterogeneous gene expression data.
• This approach enhances machine learning-based diabetes prediction by creating robust, uniform patient representations.
• Integrating multi-dataset gene expression data with domain knowledge offers a powerful strategy for improving disease prediction models.