Abstract
Background: The hemoglobin E (Hb E) trait, which is highly prevalent in Southeast Asia and may co-occur with α-thalassemia, often exhibits heterogeneous complete blood count profiles; notably, some individuals with the Hb E trait have near-normal MCV and MCH values, increasing the risk of misclassification during routine hematological screening.
Objective: This study aimed to develop and externally validate an interpretable machine-learning model to screen for Hb E trait using routine Complete Blood Count (CBC) data.
Materials and methods: We analyzed 5,787 standardized CBC records using two processing pipelines: a standard dataset and an SMOTE-NC oversampled dataset. We selected features using ReliefF and evaluated nine machinelearning algorithms. We interpreted the optimal model using SHAP values and validated it on an independent external cohort (n=625).
Results: Discrimination for Hb E trait was moderate-to-good and consistent between internal and external evaluation. Erythrocyte indices-Z-MCV, Z-MCH, and Z-RBC-were the most discriminative features. Among nine algorithms, a support vector machine gave the best combination of sensitivity and discrimination (internal AUC 0.881, sensitivity 70.66%, specificity 87.92%; external AUC 0.879, sensitivity 72.38%, specificity 87.39%). With the decision threshold tuned for screening, external sensitivity rose to approximately 91% with negative predictive value above 94%. SHAP analysis confirmed the model’s reliance on biologically plausible drivers (low MCV/MCH with a relatively preserved RBC count).
Conclusion: An explainable machine-learning model utilizing routine CBC parameters offers a viable, low-cost pre-screening tool for Hb E trait. Internal and external performance are concordant, and with local calibration and threshold tuning, the model triages patients for confirmatory testing and definitive genotyping rather than replacing them, supporting optimized diagnostic workflows in endemic regions.