Application of machine learning in screening for β-thalassemia using complete blood count (CBC) parameter

Abstract

This study is among the first in Vietnam to apply machine learning (ML) for β-thalassemia screening based solely on complete blood count (CBC) parameters. A regionally imbalanced dataset of 515 CBC samples was collected from students in Lai Châu province at Chemedic Laboratory (Hanoi) between October and December 2023. A validation set of 111 samples, including 55 β-thalassemia cases confirmed by high-performance liquid chromatography (HPLC), was analyzed using an XP-100 hematology analyzer. Supervised ML models-Decision Tree (DT), Random Forest (RF), and Logistic Regression (LR)-were developed with Python libraries (scikit-learn, TensorFlow), incorporating Synthetic Minority Oversampling Technique (SMOTE), Adaptive Synthetic Sampling (ADASYN), Principal Component Analysis (PCA), and Singular Value Decomposition (SVD) for data balancing and feature extraction. The SMOTE-PCA/SVD combinations achieved high accuracy (0.95 for DT and RF; 0.93 for LR), with ROC AUC 0.94–0.96 and F1-score ≈ 0.95. Using ADASYN with PCA/SVD improved DT accuracy to 0.97 but reduced RF to 0.85. Optimal performance occurred with a 500-sample training set, 60:40 class ratio, and test sizes of 0.05 – 0.2. These findings demonstrate that ML, particularly DT and RF, can serve as cost-effective, non-invasive screening tools for β-thalassemia in resource-limited regions of Vietnam, although further validation with larger and genetically confirmed datasets is warranted.