Machine Learning–Based Classification of Molecularly Defined α-Globin Genotypes Using Routine Hematological Parameters
International Journal of Molecular Sciences · 20.11.2026 · 27 (22) · 10450
DOI: 10.3390/ijms272210450 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Molecular Sciences, online erschienen am 20.11.2026. Verfasst von Süheyl Uçucu. Die Arbeit nennt 22 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Alpha-thalassemia is a genetically heterogeneous hemoglobinopathy caused by reduced or absent α-globin chain production. Its phenotypic presentation may overlap with other microcytic disorders, and routine hematological parameters alone may not reliably permit identification of the underlying α-globin genotype. Although molecular analysis provides definitive genotype characterization, its availability may be limited in routine screening settings. Machine learning approaches offer a potential strategy for identifying complex patterns within routinely available hematological data and may facilitate the preliminary classification of α-thalassemia molecular subtypes.This study investigated whether molecularly defined α-thalassemia groups could be classified using routinely available hematological parameters. A dataset of 599 individuals aged 18–64 years was analyzed, including 229 genotype-negative reference individuals and 370 individuals with molecularly characterized α-thalassemia genotypes. Participants were classified into five molecular groups, including individuals carrying the −α3.7 deletion, hemoglobin Constant Spring (HbCS), the --SEA deletion, and HbH/compound α-globin genotypes. Age, sex, red blood cell count (RBC), hemoglobin concentration (Hb), hematocrit (HCT), mean corpuscular volume (MCV), mean corpuscular hemoglobin (MCH), mean corpuscular hemoglobin concentration (MCHC), and red cell distribution width (RDW) were used as predictor variables. Five supervised machine learning algorithms were trained and compared, including Decision Tree, Naive Bayes, Random Forest, k-nearest neighbors, and artificial neural network models. Model development and evaluation were performed using stratified five-fold cross-validation, with data-dependent preprocessing performed within the training folds to minimize information leakage. Model performance was assessed using accuracy, balanced accuracy, precision, recall, F1 score, ROC-AUC, and PR-AUC, together with class-specific sensitivity, specificity, precision, and F1 score. Calibration was evaluated using the Brier score and log loss. Ninety-five percent confidence intervals were estimated for the principal performance measures. SHAP-based analysis was subsequently performed to assess the relative contribution of individual hematological variables to model predictions.The evaluated machine learning models demonstrated meaningful discriminatory performance for classifying molecularly defined α-thalassemia groups using routine hematological parameters. Naive Bayes achieved the highest overall accuracy (78.8%), while Random Forest showed the highest macro-ROC-AUC (0.916) and macro-PR-AUC (0.743), together with the most favorable comparative calibration performance. SHAP-based analysis identified MCH, MCV, and RDW as the most influential predictors of model decisions. Classification performance varied across molecular subgroups, with lower sensitivity observed for the Hb Constant Spring group. Routine hematological parameters contain informative patterns associated with molecularly defined α-thalassemia groups. In this cohort, machine learning models were able to distinguish individuals with α-thalassemia from genotype-negative reference individuals and to discriminate among different molecular α-globin gene patterns, although performance varied across subgroups. These findings support the potential use of machine learning as a complementary, non-invasive, and readily accessible approach for the preliminary classification and screening of α-thalassemia. However, molecular analysis remains essential for definitive genotype characterization, and external validation in independent and ethnically diverse cohorts is required before clinical implementation.
Zitieren
APA
Uçucu, S. (2026). Machine Learning–Based Classification of Molecularly Defined α-Globin Genotypes Using Routine Hematological Parameters. International Journal of Molecular Sciences, 27(22), 10450. https://doi.org/10.3390/ijms272210450
Vancouver
Uçucu S. Machine Learning–Based Classification of Molecularly Defined α-Globin Genotypes Using Routine Hematological Parameters. International Journal of Molecular Sciences. 2026;27(22):10450. doi:10.3390/ijms272210450
BibTeX
@article{Uucu2026,
author = {Uçucu, Süheyl},
title = {Machine Learning–Based Classification of Molecularly Defined α-Globin Genotypes Using Routine Hematological Parameters},
journal = {International Journal of Molecular Sciences},
year = {2026},
volume = {27},
number = {22},
pages = {10450},
doi = {10.3390/ijms272210450}
}
EUVIMED