Federated Semi-Supervised GNN for Privacy-Preserving GDM Prediction
A new machine learning framework, FedTGNN-SS, addresses two key challenges in predicting Gestational Diabetes Mellitus (GDM) from electronic health records: label scarcity and data privacy. The framework operates in a federated setting where each hospital trains a local graph neural network (GNN) on a k-nearest-neighbor patient similarity graph, without sharing raw patient data. To leverage unlabeled data, it employs prototype-guided pseudo-labeling with neighborhood agreement, adaptive graph refinement, and clinical-aware consistency. The approach aims to improve early risk stratification for GDM, a high-prevalence pregnancy complication, while preserving patient privacy. The paper is published on arXiv under identifier 2605.01810.
Key facts
- FedTGNN-SS is a federated semi-supervised framework for tabular EHR data.
- It uses local k-NN patient similarity graphs and topology-adaptive GNN encoders.
- Prototype-guided pseudo-labeling with neighborhood agreement exploits unlabeled records.
- Adaptive graph refinement periodically updates the k-NN graph using learned embeddings.
- Clinical-aware consistency is incorporated to enhance learning.
- The framework addresses label scarcity and data privacy in GDM prediction.
- GDM is a high-prevalence pregnancy complication requiring early risk stratification.
- The paper is available on arXiv with ID 2605.01810.
Entities
Institutions
- arXiv