TrustFedKG-Health: Explainability-Aware Personalized Federated Knowledge Graph Neural Network with Uncertainty-Guided Clinical Reasoning
Abstract
Abstract During preliminary experiments with federated graph learning on MIMIC-IV, we observed a consistent failure mode: hospitals with large patient volumes systematically dominated model aggregation even when their EHR records contained substantially higher rates of missing lab values and inconsistent diagnostic coding. A hospital contributing 30% of training samples but with 18% missing lab values outweighed a smaller, well-curated site in every round of standard FedAvg training. This quality-blindness motivated the core design of TrustFedKG-Health. A second observation drove the knowledge graph component: patient similarity graphs constructed purely from feature statistics missed clinically established disease-drug-symptom co-occurrences-two patients sharing a metformin-T2DM-polyuria triple were treated as unrelated if their raw feature vectors differed. Addressing both problems simultaneously, without requiring hospitals to share patient data, defined the design space for this work. We present TrustFedKG-Health, a federated knowledge graph neural network for disease risk prediction. It addresses these failures through five components. First, XFedGAT, a quality-aware aggregation algorithm weighting hospitals by dataset quality, prediction confidence, and explanation stability. Second, a Medical Knowledge Graph (MKG) layer encoding UMLS disease-drug-symptom triples as additional graph edges. Third, Monte Carlo dropout for calibrated predictive uncertainty. Fourth, dual…
Read the paper
Related studies
- Metformin as an adjuvant in breast cancer treatment
- Metformin Lysosomal Targeting: A Novel Aspect to Be Investigated for Metformin Repurposing in Neurodegenerative Diseases?
- Aberrant HSF1 signaling activation underlies metformin amelioration of myocardial infarction in mice
- Metformin induces ZFP36 by mTORC1 inhibition in cervical cancer-derived cell lines
- Protein Interaction Network for Identifying Vascular Response of Metformin (Oral Antidiabetic)
- Preventing Long COVID With Metformin
Go deeper in Evidence Hub
Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.