<![CDATA[Anemia remains a major global health concern that impairs oxygen transport and contributes to fatigue, cognitive decline, reduced productivity, and severe clinical complications. Although machine learning has shown promise for automated anemia detection, multiclass classification remains challenging due to class imbalance, overlapping hematological characteristics, and limited model interpretability. This study proposes an explainable hybrid framework integrating Extreme Gradient Boosting (XGBoost), SHapley Additive exPlanations (SHAP), and Fuzzy Logic to improve anemia risk classification and clinical decision support. The publicly available SKILICARSLAN dataset containing 15,300 anonymized patient records across five anemia-related classes was utilized. Seven hematological parameters, namely hemoglobin (HGB), mean corpuscular volume (MCV), mean corpuscular hemoglobin (MCH), mean corpuscular hemoglobin concentration (MCHC), red blood cell count (RBC), hematocrit (HCT), and ferritin, were employed as predictive features. The workflow comprised data auditing, stratified train–test splitting, Synthetic Minority Oversampling Technique (SMOTE), hyperparameter optimization, multiclass XGBoost modeling, SHAP-based explainability analysis, and fuzzy risk interpretation. Experimental results demonstrated 82.94% accuracy, 87.27% weighted precision, 82.94% weighted recall, and 84.78% weighted F1-score, with a mean cross-validation F1-score of 87.00%. The model further achieved a macro-average ROC–AUC of 0.81 and a weighted-average ROC–AUC of 0.90, indicating robust discriminative performance despite class imbalance. SHAP analysis identified HGB, ferritin, and RBC-related variables as the most influential predictors. Moreover, the fuzzy logic layer enhanced interpretability by translating model outputs into clinically meaningful risk levels. These findings demonstrate the potential of explainable hybrid intelligence for transparent and reliable anemia screening and decision-support applications.]]>
