Predicting anemia risk in Iranian adults without diagnostic blood indices: an interpretable machine learning approach integrating dietary

Predicting anemia risk in Iranian adults without diagnostic blood indices: an interpretable machine learning approach integrating dietary, clinical, and non-diagnostic laboratory patterns. Download PDF Abstract Anemia remains a major public-health burden, disproportionately affecting adults of reproductive age. In the original submission, an interpretable machine-learning framework trained to predict anemia status in Iranian adults achieved a near-perfect AUC (≈ 1.00) using XGBoost; closer methodological scrutiny of this result identified two critical flaws that undermine its validity as a measure of genuine predictive power: (i) hemoglobin-defining hematological indices (HGB, HCT, RBC, MCV, MCH) were included as predictors, creating circular prediction, and (ii) feature selection was performed on the full dataset before train/test splitting, causing data leakage.

This revised analysis removes both flaws: diagnosis-defining indices were excluded from the predictor set, and the entire pipeline — feature selection, SMOTE oversampling, and hyperparameter tuning — was rebuilt inside a leakage-free pipeline confined strictly to each training fold. Twelve supervised algorithms were compared on an independent, held-out test set ( n = 641 of 3,203 records). Logistic Regression achieved the best discrimination (AUC = 0.8625, 95% bootstrap CI 0.8168–0.9053; Recall = 0.7738, 95% CI 0.6854–0.8589; F1 = 0.5350, 95% CI 0.4561–0.6081; Accuracy = 0.8237), clearly outperforming a simple demographic baseline (AUC = 0.5284).

The drop in AUC from ≈ 1.00 to 0.86 after removing the diagnosis-defining indices empirically confirms that the original near-perfect performance was an artifact of circular prediction and data leakage. SHAP analysis of the corrected model shows that, alongside dietary and lifestyle variables (dairy-product intake, carbohydrate intake, household-cleaning activity), several non-diagnostic hematologic markers (RDW-CV, platelet count, WBC count, PDW) remain influential — an important nuance discussed below, since these still require venipuncture even though they are not part of the anemia diagnostic criterion. The corrected framework offers a more honest, generalizable, and clinically defensible foundation for anemia risk screening in Iranian adults.

Medical Disclaimer: This article is for informational purposes only and does not constitute medical advice. Always consult with qualified healthcare professionals for medical decisions and treatment options.