CS-GIN: A cosine similarity-based graph isomorphism network for multi-class anemia classification

Accurate anemia diagnosis remains challenging due to the heterogeneous characteristics of anemia subtypes and the complex interactions among hematological biomarkers. Conventional machine learning approaches generally treat patient records as independent observations, limiting their ability to exploit clinically meaningful relationships among patients with similar hematological profiles. To address this limitation, this study proposes a Cosine Similarity-based Graph Isomorphism Network (CS-GIN) for multi-class anemia classification. A hematological dataset consisting of 513 patient records was transformed into a patient similarity graph using cosine similarity, where patients were represented as graph nodes and similarity relationships were represented as graph edges. A parameter sensitivity analysis was first conducted to determine the optimal cosine similarity threshold, followed by comparisons with Euclidean distance- and K-nearest neighbor (KNN)-based graph construction strategies. The resulting graphs were evaluated using Graph Convolutional Network (GCN), Graph Attention Network (GAT), conventional Graph Isomorphism Network (GIN), and the proposed CS-GIN under stratified five-fold cross-validation. Experimental results demonstrated that CS-GIN achieved the best performance among all graph neural network models, obtaining an average accuracy of 90.24%, outperforming GCN (84.39%), GAT (84.59%), and conventional GIN (83.23%). Statistical significance was confirmed using the Friedman and Nemenyi tests. Explainability analyses using SHAP and GNNExplainer consistently identified Ferritin, HbA2, RET-He, and MCV as the most influential biomarkers, while graph topology and embedding analyses demonstrated that the learned representations effectively preserved meaningful patient similarity relationships. These findings demonstrate that similarity-aware graph construction substantially enhances graph representation learning and provides an effective, interpretable framework for graph-based multi-class anemia classification.