The study explores the use of Graph Neural Networks (GNNs) for disease prediction in various medical scenarios. It proposes a methodology that transforms tabular patient data into graph-structured representations, capturing intricate linkages and patterns inherent in healthcare data. The methodology includes five essential stages: data preprocessing, graph generation, node embedding and feature extraction using GNNs, feature integration, and machine learning classification. During the graph generation step, similarity metrics like Gaussian, Jaccard, Cosine, and Spearman correlations are used to generate patient graphs that represent different aspects of patient similarities. Link prediction is used as an unsupervised learning objective to train GNN models, including Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), and GraphSAGE, to acquire significant node embeddings. The latent features obtained by GNNs are then combined with baseline preprocessed features, resulting in an enhanced feature set that incorporates both node-level and graph-level information. Machine learning classifiers such as Support Vector Machines (SVM) and Random Forest are trained using this enriched feature set. The empirical findings show that GNN-based approaches consistently surpass earlier state-of-the-art methods in all three disease categories. Graph Attention Network (GAT) shows outstanding performance in stroke prediction. The integration of GNN information with conventional classifiers leads to synergistic effects, producing highly accurate and robust prediction models. This work significantly enhances the medical informatics domain by demonstrating the remarkable capabilities of graph-based deep learning in disease prediction. It suggests that GNN-based methods can significantly increase the precision and dependability of disease prediction models, leading to earlier detection, more tailored treatment approaches, and better patient outcomes.
