Infant Low Birth Weight Prediction Using Graph Embedding Features.

Area of Science:

• Medical Informatics
• Machine Learning
• Public Health

Background:

• Low Birth Weight (LBW) is a significant global health issue with short- and long-term consequences for infants and mothers.
• Accurate prenatal prediction of infant weight is essential for identifying risk factors and mitigating infant morbidity and mortality.
• Current Machine Learning (ML) models for LBW prediction require performance enhancement for real-world clinical adoption, as they often overlook structural data patterns.

Purpose of the Study:

• To enhance the performance of LBW classification by developing a novel approach using knowledge graphs.
• To capture complex relationships and structural information within patient data that traditional ML models may miss.

Main Methods:

• Transformed tabular patient data into a knowledge graph structure.
• Extracted various node features, including node embeddings (node2vec), node degree, node similarity, and nearest neighbors.
• Evaluated ML models on the original dataset, graph-derived features, and a combined feature set using data from a 3453-patient cohort in the UAE.

Main Results:

• The proposed knowledge graph-based method achieved a significant improvement in LBW classification performance.
• The model reached an Area Under the Curve (AUC) of 0.834, representing over a 6% improvement compared to models using only original risk factors.
• The study demonstrated the clinical relevance of the developed model for potential integration into clinical settings.

Conclusions:

• Knowledge graph transformation of patient data offers a promising avenue for improving LBW prediction accuracy.
• The enhanced predictive performance holds potential for better clinical decision-making and improved infant outcomes.
• The developed model's clinical relevance supports its potential adaptation in real-world healthcare scenarios.