Machine Learning Classifiers for Differentiation between Iron Deficiency Anaemia and Beta Thalassemia Trait: comparative study

Abstract

This research compares various machine learning classifiers to differentiate between Iron Deficiency Anaemia (IDA) and Beta Thalassemia trait (BTT). The goal is to identify the most suitable classifiers for handling complex and intertwined medical data, systematically evaluating twelve machine learning (ML) algorithms on a locally curated dataset comprising 2,160 participants (1,080 diagnosed with IDA and 1,040 with BTT conditions). The models being assessed include classical classifiers—Logistic Regression, Decision Tree, Random Forest, Support Vector Machine (SVM), Multilayer Perceptron (MLP), Linear Discriminant Analysis (LDA), k-Nearest Neighbors (KNN), and Naive Bayes—as well as ensemble-based methods: Gradient Boosting, AdaBoost, XGBoost, and CatBoost. Comprehensive pre-processing steps were applied, including outlier data deletion, imputation of missing values, class balancing, and feature scaling. An 80:20 train-test split was employed, with performance validated using 5-fold cross-validation to mitigate overfitting risks. Results indicate that SVM achieved the highest accuracy (97.0%) with an AUC of 99.1%, sensitivity of 95.9%, specificity of 98.7%, and the fastest execution time (0.37 seconds). AdaBoost and MLP followed closely, attaining accuracies of 96.9% and 96.8%, respectively. All models demonstrated high robustness, with F1-scores exceeding 96.8%. SVM provided the best trade-off between diagnostic performance and computational efficiency. These results emphasize the promise of enhanced machine learning models, especially SVM, as dependable and economical diagnostic assistance tools for IDA_BTT differentiation in clinical settings, providing a substitute for costly laboratory tests while maintaining diagnostic precision. The proposed framework underscores the feasibility of deploying ML techniques in haematological diagnostics based on routine clinical data.