Machine Learning-Based Multiclass Classification of Hb E Carrier Status and Thalassemia-Related Hematological Profiles Using Routine Complete Blood Count Parameters.

Routine complete blood count (CBC) measurements are widely available and may support the initial assessment of thalassemia-related hematological conditions. However, overlapping hematological profiles and class imbalance can complicate differentiation among groups. This study evaluated seven supervised machine learning classifiers for multiclass classification of normal hematological profiles, suspected α-thalassemia profiles, Hb E carrier-related profiles, and other thalassemia-related/hemoglobinopathic profiles using a publicly available CBC dataset. An unresampled baseline was compared with four synthetic oversampling strategies: SMOTE, Borderline-SMOTE, SVM-SMOTE, and KMeans-SMOTE. Model development and hyperparameter optimization were performed within five-fold stratified cross-validation, with model selection based on cross-validated Matthews correlation coefficient (MCC). The untouched internal hold-out test set was reserved for final evaluation. CatBoost without resampling yielded the highest cross-validated MCC (0.6378). On the internal hold-out test set, it achieved an accuracy of 0.7950, an MCC of 0.6537, Cohen's kappa of 0.6499, and micro- and macro-average ROC-AUC values of 0.9554 and 0.9227, respectively. However, class-specific sensitivity was heterogeneous, ranging from 0.4028 for suspected α-thalassemia to 0.9767 for the normal hematological profile. Synthetic oversampling did not consistently increase overall CV MCC, although sensitivity improved for some minority classes. These findings characterize the performance and limitations of CBC-based multiclass classification and indicate the need for further external and clinically oriented validation before screening or referral applications can be considered.