Knowledge Graph–Driven Laboratory Intelligence for Anemia Diagnosis Using Hematological Profiles

Accurate anemia characterization from laboratory data remains challenging due to complex physiological interactions and class imbalance among diagnostic categories. Therefore, this study aimed to propose a knowledge graph–based framework that transformed hematological measurements into relational structures connecting laboratory parameters, physiological mechanisms, and diagnostic outcomes. Knowledge Graph Embeddings (KGE) were also used to learn structured representations, while the Synthetic Minority Oversampling Technique with Multiresolution Sampling (SMOTE-MRS) was incorporated to mitigate imbalance and preserve local distribution patterns. Patient-level embeddings derived from the relational graph were evaluated using four machine learning classifiers, namely Random Forest, Extra Trees, Support Vector Machine, and XGBoost. The experimental results showed that relational modeling achieved classification accuracies of 0.807 using complete blood count (CBC) parameters and 0.815 when extended hematology features were included. After applying SMOTE-MRS, performance increased substantially to 0.980 and 0.986 for CBC data and CBC combined with extended features. Embedding space visualization suggested clearer class separability after multiresolution sampling. Statistical evaluation using the Friedman test reported no significant performance differences among classifiers without SMOTE-MRS (p = 0.233 and p = 0.392). Meanwhile, significant differences were obtained after SMOTE-MRS balancing (p = 0.018 and p = 0.016), showing improved exploitation of structured embeddings under balanced data conditions. Comparative analysis with previous knowledge graph–based studies showed competitive predictive performance while preserving clinically meaningful relationships. Integrating relational modeling with imbalance-aware sampling provided a promising pathway toward robust and interpretable laboratory intelligence systems for anemia screening and subtype differentiation. • Knowledge graph integrates hematology data for anemia diagnosis • Hierarchical graph models lab markers and disease mechanisms • SMOTE-MRS preserves minority anemia phenotypes during training • Graph embedding improves classification robustness across models • Extended hematology profiles enhance diagnostic separability