A Graph Neural Network-Based Clinical Decision Support System for Personalized Cancer Prognosis Integrating Genomic, Radiological, and Histopathological Data

Abstract

Background: Proper prognosis of adolescent cancer is still a difficult issue as genomic, radiological, histopathological and clinical data are heterogenous and hard to coordinate and construct. Purpose: The suggested research is a Graph Neural Network (GNN)-based Clinical Decision Support System to individually prognose cancer through the integration of multimodal data. Procedures: An artificial dataset consisting of 8,000 patients, 250 genomic biomarkers, 180 radiological features, 220 histopathology features and 35 clinical variables were simulated as a patient similarity graph and tested on an 80:10:10 train-validation-test split. Findings: The proposed multimodal GNN model had an accuracy of 96.4% and an F1-score of 95.7% and a C-index of 0.947, and the AUC of 0.983, compared to traditional machine learning and deep learning models, and offers explainable predictions with SHAP and GNNExplainer with an inference latency of 23.8 ms.