Beyond prediction: explainable machine learning for ferritin estimation from complete blood count

BACKGROUND: Serum ferritin is a primary marker of body iron stores; however, its diagnostic utility is limited by cost and non-specific elevation during inflammatory states. This study aimed to develop and validate explainable machine learning (ML) models to estimate ferritin status using only widely available and cost-effective complete blood count (CBC) and demographic parameters. METHODS: This prospective study analyzed 12,450 adult laboratory records. To ensure ferritin reflected true iron status, cases with inflammation (elevated white blood cell count or erythrocyte sedimentation rate) were excluded. A balanced dataset of 7,200 samples was used to train five supervised ML algorithms, including Logistic Regression, XGBoost, and Artificial Neural Network (ANN), for both binary classification and regression tasks. Model interpretability was assessed using SHapley Additive Explanations (SHAP) to evaluate global feature importance. RESULTS: The ANN model achieved the best classification performance, with an area under the curve of 0.837 and an F1-score of 0.760 in discriminating low ferritin levels (< 30 ng/mL) from normal values. Although the regression model demonstrated a moderate global R² of 0.32, it showed higher predictive accuracy in lower ferritin ranges critical for diagnosing iron deficiency. SHAP analysis revealed Sex, Mean Corpuscular Hemoglobin, and Red Cell Distribution Width as the most influential predictors, consistent with iron-restricted erythropoiesis. CONCLUSION: ML models based solely on routine hemogram data provide a reliable, zero-cost screening approach for iron deficiency. The integration of explainable AI enhances clinical credibility and supports implementation as an accessible decision support tool.