GDR-GNN: A Graph Neural Network Framework for Explainable Hereditary Genetic Disorder Risk Prediction

Abstract

Accurate prediction of risk of hereditary genetic disorder is necessary for early intervention and clinical decision making. This study proposes a novel graph based learning model, considering clinical and genetic tabular data to be converted into a relational k nearest neighbor graph. Genetic Disorder Risk Graph Neural Network (GDR-GNN), to understand the inter-sample dependency. Automatic hyper-parameter optimization Optuna to perform to search the parameter space and find the optimal GNN configuration besides, on our sequence, and explainable AI techniques (SHAP and LIME) are integrated to provide both global and local interpretability mechanisms. The model is evaluated by testing it on a clinically validated dataset of 100 families and comparing it to the best- performing machine learning classifiers, including SVM, KNN, Random Forest, SVM, AdaBoost and Histogram based gradient boosting. The experiment outcomes show that the proposed GDR-GNN can achieve better predictive accuracy than the traditional models with the accuracy of 97.67 % and Cohen kappa of 0.9651, which is superior to the traditional models and provides greater transparency. The XAI analyses indicate that the model is based around clinical and genetic attributes such as parental carrier status, consanguinity, and the range of specific gene values, which have a central role in the model, and supports its clinical reliability. These findings highlight the potential of graph-based modelling and explainability-driven methods in order to advancing hereditary risk prediction for real world clinical use.