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

• DOI: 10.1186/s12911-024-02649-2
• Publication Date: 2024-09-02T09:03:05Z
• Authors: Tuong Minh Nguyen; Kim Leng Poh; Shu-Ling Chong; Jan Hau Lee
• Keywords: Representation

Modeling patient data, particularly electronic health records (EHR), is one of the major focuses machine learning studies in healthcare, as these provide clinicians with valuable information that can potentially assist them disease diagnosis and decision-making. In this study, we present a multi-level graph-based framework called MedMGF, which models both medical profiles extracted from EHR data their relationship network single architecture. The consist several layers embedding derived interval obtained during hospitalization, patient-patient created by measuring similarities between profiles. We also propose modification to Focal Loss (FL) function improve classification performance imbalanced datasets without need imputate data. MedMGF's was evaluated against Graphical Convolutional Network (GCN) baseline implemented Binary Cross Entropy (BCE), FL, class balancing parameter $$\alpha$$ , Synthetic Minority Oversampling Technique (SMOTE). Our proposed achieved high (AUC: 0.8098, ACC: 0.7503, SEN: 0.8750, SPE: 0.7445, NPV: 0.9923, PPV: 0.1367) on an extreme pediatric sepsis dataset (n=3,014, imbalance ratio 0.047). It yielded improvement 3.81% for AUC, 15% SEN compared GCN+ FL 0.7717, 0.8144, 0.7250, 0.8185, 0.1559, 0.9847), 5.88% AUC 22.5% GCN+FL+SMOTE 0.7510, 0.8431, 0.6500, 0.8520, 0.1688, 0.9814). showed 3.86% BCE 0.7712, 0.8133, 0.8173, 0.1551, 14.33% 27.5% comparison GCN+BCE+SMOTE 0.6665, 0.7271, 0.6000, 0.7329, 0.0941, 0.9754). When all models, MedMGF highest results, demonstrating potential healthcare applications.