Fine-grained Patient Similarity Measuring using Contrastive Graph Similarity Networks.

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Machine Learning

Background:

• Electronic Health Records (EHRs) are increasingly used for predictive analytics, driven by deep learning advancements.
• Patient representation learning from EHRs is a key area, but existing methods struggle with irregular data and patient similarity.
• Current deep learning models often overlook patient similarity, a crucial aspect of clinical reasoning.

Purpose of the Study:

• To develop a novel method for calculating patient similarity in large EHR datasets.
• To generate rich patient representations by incorporating similarity information.
• To improve downstream prediction tasks using enhanced patient representations.

Main Methods:

• A Contrastive Graph Similarity Network was developed for patient similarity calculation.
• Graph-based similarity analysis was employed to extract clinical characteristics.
• Information from similar patients was aggregated to create robust patient representations.

Main Results:

• The proposed method demonstrated effectiveness in similarity calculation among patients.
• Experimental results showed superiority over existing methods on real-world EHR data.
• The approach improved performance in vital signs imputation and ICU patient deterioration prediction.

Conclusions:

• The Contrastive Graph Similarity Network effectively addresses limitations in current EHR patient representation learning.
• Incorporating patient similarity significantly enhances predictive model performance.
• This method offers a promising approach for clinical decision support using EHR data.