Classification of α-thalassemia data using machine learning models

Area of Science:

• Medical Diagnostics
• Computational Biology
• Genetics

Background:

• Congenital hemoglobin disorders affect 7% of the global population, with over 300,000 annual alpha-thalassemia cases.
• Accurate diagnosis is challenging in low-income regions, often relying on complete blood count (CBC) tests.
• This study explores machine learning for alpha-thalassemia trait classification using CBC and gender.

Purpose of the Study:

• To develop and evaluate machine learning models for classifying alpha-thalassemia traits.
• To compare the performance of machine learning models against traditional discriminant formulae.
• To investigate the impact of grouping silent and non-carriers on classification accuracy.

Main Methods:

• Utilized a dataset of 288 individuals with suspected alpha-thalassemia from Sri Lanka.
• Employed eleven discriminant formulae and nine machine learning models for classification.
• Applied outlier removal (Mahalanobis distance) and resampling techniques (SMOTE, SMOTE-NC).

Main Results:

• Convolutional Neural Network (CNN) achieved high performance (Accuracy: 0.85, ROC-AUC: 0.95/0.93) without feature extraction.
• Ehsani formula showed moderate performance (ROC-AUC: 0.66) when grouping silent and non-carriers.
• Machine learning models significantly outperformed classical discriminant formulae.

Conclusions:

• Machine learning models offer superior performance for alpha-thalassemia classification compared to traditional methods.
• Grouping silent and non-carriers, combined with resampling, enhances ML model efficacy.
• Silent carriers remain indistinguishable from non-carriers based on current features and dataset size.