Explainable Heterogeneous Graph Neural Networks with Severity-Weighted Edges for Clinical Symptom-Disease Prediction

Automated symptom-checkers built on tabular fea-
ture vectors cannot model the relational structure inherent in
medical knowledge. We propose an Explainable Heterogeneous
Graph Convolutional Network (H-GraphConv) with Severity-
Weighted Edges for automated clinical diagnosis. Edge weights
combine empirical symptom prevalence with validated clinical
severity scores, and a Multi-Type Heterogeneous Graph incorpo-
rates Disease, Symptom, Anatomical Region, and Lab Test node
types. A Latent Feature Attribution module delivers mathemat-
ically exact, per-symptom explanations without approximation,
while a Graph-Retrieval Augmented Generation (GraphRAG)
pipeline powered by Google Gemini generates clinician-facing
diagnostic narratives grounded in GNN-derived graph facts. On a
4,920-record dataset covering 41 diseases and 131 symptoms, the
model attains 95.6% link-prediction accuracy, 95.7% F1-score,
and AUC-ROC of 0.9412. Under simulated symptom-dropout
noise of 0–30%, Top-5 disease-ranking accuracy reaches ≈33%,
which is 2.7× the random baseline.
Index Terms—Graph Neural Networks, Explainable AI, Latent
Feature Attribution, Link Prediction, Heterogeneous Graphs,
Medical Diagnosis, Symptom Checker, Clinical Decision Support,
GraphRAG, PyTorch Geometric.