<![CDATA[Anemia remains a serious global health problem, yet its classification from hematological data faces issues of class imbalance and clinical interpretability. In this study, we propose an explainable Graph Neural Network (GCN) framework integrated with Synthetic Minority Over-sampling Technique (SMOTE) and GNNExplainer for anemia classification. Addressing the limitations of the previous global similarity threshold, we construct a localized k-nearest neighbors (k-NN) graph with k=5 and apply clinical-range filtering to remove physiologically impossible outliers. Evaluation on 136 test samples shows that the proposed GCN model achieves an accuracy of 91.91% and a recall of 82.76% for the anemia class, which is highly competitive with SVM RBF (93.38% accuracy) and Naive Bayes (91.91% accuracy). McNemar significance tests confirm that the GCN model performs comparably to the baselines (p >= 0.72) while offering model interpretability. Using GNNExplainer, the parameters MCH, Ht, HB, and MCHC were identified as the most dominant features, providing clinicians with valuable transparency. This framework demonstrates the clinical utility of explainable GNNs in medical diagnostics.]]>
Copyrights © 2026
