Anemia screening in primary care benefits substantially from models that rely exclusively on standard complete blood count (CBC) measurements. This study evaluates six machine learning classifiers for binary anemia detection and tests whether a transparent linear model can match complex ensembles. We analyzed a public clinical dataset of 1,004 records collected from Medical Centre Hospital, Chattogram, Bangladesh. After removing one incomplete entry and 27 exact duplicates, the curated cohort contained 976 unique cases with eight routine predictors: age, sex, hemoglobin (HGB), red blood cell count (RBC), packed cell volume (PCV), mean corpuscular volume (MCV), mean corpuscular hemoglobin (MCH), and mean corpuscular hemoglobin concentration (MCHC). Classifiers were compared using stratified 10-fold cross-validation on the training set (n = 780). A prespecified parsimony rule selected Logistic Regression as the primary model when its mean cross-validated ROC-AUC remained within 0.02 of the leading candidates. On the held-out test partition (n = 196), Logistic Regression attained an accuracy of 0.9796, an F1-score of 0.9753, and a ROC-AUC of 0.9978. Calibration analysis revealed a slope of 2.0274 and an intercept of 0.2893, indicating that raw probability outputs require local recalibration before clinical deployment. Global SHAP explanations confirmed PCV, RBC, and HGB as primary predictors, aligning with local LIME surrogate explanations. Single-feature removals produced negligible performance drops for secondary indices but notable sensitivity for RBC and HGB. Subgroup evaluations showed consistent accuracy and sensitivity across the evaluated sex and age strata, although the relatively small subgroup sizes limit conclusions about broader demographic generalizability.

Anemia screening,complete blood count,explainable artificial intelligence,logistic regression,probability calibration,subgroup robustness