Reproduce “Knowledge Graph Representations to enhance Intensive Care Time-Series Predictions” on OpenResearch

About this paper

Samyak Jain, Manuel Burger, Gunnar Rätsch, Rita Kuznetsova

A new method combines knowledge graph representations from clinical ontologies with patient vital signs and clinical notes to enhance deep learning models for time-series predictions in Intensive Care Units. This approach yields higher accuracy on MIMIC-III benchmarks and increased robustness to missing data, while also offering interpretability of key medical concepts influencing predictions.

Key insights

• Explicitly integrating structured medical knowledge graphs from clinical ontologies into deep learning models can substantially boost predictive performance for ICU time-series tasks.
• Multimodal graph-structured representations, combining vital signs, text, and knowledge, lead to increased robustness in predictions, particularly when dealing with missing clinical data.
• Leveraging GNN attention scores allows for the identification of influential medical concepts, enhancing the interpretability of model predictions for clinicians.

Summary by alphaXiv.