Graph neural network modelling as a potentially effective method for predicting and analyzing procedures based on...

Area of Science:

• Health Informatics
• Machine Learning
• Network Science

Background:

• Healthcare aims to improve patient care quality and economic efficiency.
• Electronic Health Records (EHRs) can identify disease and therapy patterns for best practice guidelines.
• Recommender systems can be implemented based on identified patterns, linking procedure volume to model quality.

Purpose of the Study:

• To develop a novel method for clustering similar patients using graph-data representation.
• To leverage Graph Neural Networks (GNNs) for analyzing patient graphs to recommend appropriate medical procedures.
• To address limitations of existing machine learning methods that ignore population structure and patient similarity.

Main Methods:

• Developed a graph-data representation to cluster similar patients based on shared patterns.
• Constructed a patient graph linking patients with similar diagnoses and typologies.
• Utilized Graph Neural Networks (GNNs) to analyze the patient graph and identify relevant medical procedures.

Main Results:

• Successfully constructed patient graphs using basic information, diagnoses, and trained GNN models.
• GNN models demonstrated superior performance compared to baseline models, with an average F1 score improvement of 6.48%.
• GNNs enabled effective clustering analysis for identifying specific therapeutic clusters related to diagnosis combinations.

Conclusions:

• GNN models show promise for modeling diagnosis distribution and identifying patients with similar phenotypes based on comorbidities.
• Challenges remain in graph construction, including potential biases and dependency on diagnostic data quality.
• Further research is needed to enhance patient embedding in graph structures for improved future applications.