Med-MGF: multi-level graph-based framework for handling medical data imbalance and representation.

Area of Science:

• Machine Learning in Healthcare
• Graph Neural Networks
• Data Mining

Background:

• Electronic Health Records (EHR) are crucial for clinical decision-making.
• Modeling patient data from EHRs is a key focus in healthcare machine learning.
• EHR data offers valuable insights for disease diagnosis and treatment.

Purpose of the Study:

• To present MedMGF, a multi-level graph-based framework for EHR data analysis.
• To model patient medical profiles and their relationship network in a unified architecture.
• To improve classification performance on imbalanced datasets using a modified Focal Loss function.

Main Methods:

• Developed MedMGF, a graph-based framework integrating patient profiles and relationship networks.
• Created patient-patient networks by measuring similarity between embedded medical profiles.
• Proposed a modified Focal Loss function to enhance classification accuracy on imbalanced data.
• Evaluated MedMGF against baseline Graphical Convolutional Network (GCN) models using various loss functions and oversampling techniques.

Main Results:

• MedMGF achieved high classification performance (AUC: 0.8098, SEN: 0.8750) on an imbalanced pediatric sepsis dataset (imbalance ratio of 0.047).
• Demonstrated significant improvements in AUC (up to 14.33%) and Sensitivity (up to 27.5%) compared to baseline models.
• Outperformed GCN models using Binary Cross Entropy (BCE), Focal Loss (FL), and Synthetic Minority Oversampling Technique (SMOTE).

Conclusions:

• MedMGF achieved superior Sensitivity (SEN) and Area Under the Curve (AUC) compared to all evaluated baseline models.
• The framework shows significant potential for various healthcare applications, particularly in analyzing complex EHR data.
• The modified Focal Loss function effectively addresses challenges posed by imbalanced datasets without imputation.