Interpretable graph hyperinsulinemia detection model with orthogonal array optimization

Early detection of hyperinsulinemia in adolescents offers a promising opportunity to significantly reduce or even eliminate the risk of developing type 2 diabetes, cardiovascular diseases, and other chronic non-infectious conditions. Graph Neural Networks (GNNs), though complex, can provide precise forecasts, which helps narrow the gap between expected and actual diagnostic classifications in medical tasks such as detecting hyperinsulinemia, thus minimizing associated health risks. However, defining optimal hyperparameter configurations for GNNs remains a challenge. In this paper, we compare state-of-the-art models, including Long Short-Term Memory (LSTM), Prior Structural Information-Graph Neural Network (PSI-GNN), and GraphSage (Graph Sample and AggreGatE). Our goal is to identify the best model for creating a balance between computational efficiency and prediction accuracy within a dataset collected nationally by the Healthcare Center in Serbia. We optimize performance through experiments with various hyperparameter settings and apply SHAP (SHapley Additive exPlanations) for a more precise determination of feature relevance, as evidenced by the reduced norm of activation vectors when compared with other centrality measures. Our findings reveal that the Taguchi orthogonal array optimization method, when applied to PSI-GNN, is the most computationally efficient. It yields superior performance metrics: an accuracy of 93.5%, an AUC of 97.4%, a precision of 92.2%, a recall of 96.5%, and an F1 score of 96.3%, with a running time of 273.2 seconds for 8 runs.