Node embedding-based graph autoencoder outlier detection for adverse pregnancy outcomes.

Area of Science:

• Medical Informatics
• Machine Learning
• Public Health

Background:

• Adverse pregnancy outcomes, including low birth weight (LBW) and preterm birth (PTB), pose significant risks to maternal and infant health.
• Early prediction is crucial for effective prevention strategies.
• Traditional machine learning models struggle with imbalanced medical data and complex relationships.

Purpose of the Study:

• To develop and evaluate a novel node embedding-based graph outlier detection algorithm for predicting adverse pregnancy outcomes.
• To address limitations of existing machine learning approaches in handling imbalanced datasets and intricate data structures.

Main Methods:

• Constructed a knowledge graph from a curated Emirati population dataset.
• Employed two node embedding algorithms and a graph autoencoder (GAE).
• Identified adverse pregnancy outcomes as outliers based on GAE reconstruction difficulty.

Main Results:

• Incorporating node embeddings into the GAE model significantly enhanced prediction performance.
• Achieved a 12% higher AUC-ROC compared to traditional GAE models.
• Demonstrated improved prediction accuracy for LBW, PTB, and very PTB datasets.

Conclusions:

• Node embedding and graph outlier detection are effective strategies for improving adverse pregnancy outcome prediction.
• This approach shows promise for well-curated population datasets.
• Highlights the potential of advanced machine learning techniques in perinatal health.