Evaluation of Graph-Based Algorithms for Early Detection of In-Hospital Mortality.

Area of Science:

• Clinical Informatics
• Machine Learning in Healthcare
• Patient Safety Research

Background:

• Preventing health-related adverse events (HAEs) is crucial for enhancing patient care quality.
• Machine learning (ML) tools offer potential for predicting HAEs, including in-hospital mortality.
• Graph neural networks (GNNs) are underutilized for analyzing complex, interconnected clinical data from clinical data warehouses (CDWs).

Purpose of the Study:

• To evaluate the performance of two GNN models for predicting in-hospital death.
• To benchmark GNN models against established methods using real-world clinical data.
• To explore the potential of GNNs for optimizing predictions and improving explainability in healthcare.

Main Methods:

• Utilized two distinct GNN models for predictive analysis.
• Benchmarked GNN performance against strong baseline methods.
• Employed data from a clinical data warehouse (CDW) representing complex, interconnected patient information.

Main Results:

• GNN models demonstrated performance comparable to the best existing benchmarks for in-hospital death prediction.
• The study confirmed the viability of GNNs in analyzing complex clinical hospital data.
• Identified opportunities for GNN model optimization and enhanced prediction explainability.

Conclusions:

• GNNs represent a promising, yet underexplored, approach for predicting in-hospital mortality.
• The graph structure inherent in clinical data offers unique advantages for ML-driven healthcare analytics.
• Further research into GNNs can lead to significant advancements in patient safety and care quality.