Evaluating the Predictive Features of Person-Centric Knowledge Graph Embeddings: Unfolding Ablation Studies

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Data Science

Background:

• Developing predictive models with complex biomedical data is challenging due to data heterogeneity, standardization issues, and sparseness.
• Previous work introduced a person-centric ontology and a representation learning framework for extracting person-centric knowledge graphs (PKGs) and training Graph Neural Networks (GNNs).

Purpose of the Study:

• To systematically examine the results of GNN models trained with both structured and unstructured information from the MIMIC-III dataset.
• To demonstrate the robustness of the proposed approach in identifying predictive features for readmission prediction.

Main Methods:

• Utilized a person-centric ontology and representation learning to create PKGs.
• Trained Graph Neural Networks (GNNs) on PKGs derived from structured and unstructured data in the MIMIC-III dataset.
• Conducted ablation studies on clinical, demographic, and social data to assess feature importance.

Main Results:

• The GNN models demonstrated robustness in identifying predictive features within PKGs.
• Ablation studies confirmed the significance of various data types in improving prediction accuracy.
• The approach successfully identified key features for the task of readmission prediction.

Conclusions:

• The proposed systematic approach enhances the interpretability and robustness of GNN models for biomedical prediction.
• Person-centric knowledge graphs combined with GNNs offer a powerful framework for leveraging complex patient data.
• This methodology holds promise for improving patient readmission prediction and personalized healthcare.