PubMedThe Journal of the Association of Physicians of India2026-10-22
Feasibility of a Model-based HbA2 Prediction for Low-cost Screening of β-thalassemia Trait Using Blood-derived Indices: Opportunities and Challenges.
Peraka Ramadevi R, M Pradeep Kumar Reddy PKR, S Vanajakshi V, Avadhanam Sumedha S et al.
β-thalassemia trait (β-TT) is a common genetic condition with significant public health implications, particularly in regions with high prevalence. Diagnosis conventionally relies on HbA2 quantification using high-performance liquid chromatography or Hemoglobin electrophoresis, which are costly and inaccessible in resource-limited settings. Developing low-cost screening/triage tools is therefore essential.
A retrospective single-center cross-sectional study was conducted at the Department of Pathology, Apollo Institute of Medical Sciences & Research, Hyderabad, between February 2023 and December 2025. Data from 129 patients with complete blood counts, iron profiles, and HbA2 measurements were analyzed. LASSO method was applied to identify salient predictors of HbA2 from demographic, hematological, and iron parameters. Two ordinary least squares regression models (OLS) were constructed: one excluding iron parameters (model 1) and another incorporating iron parameters (model 2). LASSO-then-OLS approach was used for both 5-fold cross-validation and final model development. ROC analysis was also done.
Of the 129 patients, 15 (11.6%) had β-TT (HbA2 >3.5%). The 5-fold cross-validation found lower root mean squared error (RMSE) and superior variable selection stability with an iron parameter-inclusive model. Salient predictors in the final model were age, RBC count, MCV, and RDW-CV in model 1, while model 2 additionally included serum ferritin, transferrin saturation, and total iron binding capacity (TIBC). Age and RBC count were positively associated with HbA2, while MCV, RDW-CV, serum ferritin, and TIBC were negatively associated. Model 1 explained 40% of HbA2 variance, while model 2 explained 47% variance. Model imprecision was more pronounced for β-TT cases. ROC analysis showed an area under the curve (AUC) value of 0.868 (0.766-0.969) for model 1 and 0.904 (0.829-0.98) for model 2. Optimal cutoff (3.31% from model 1 and 3.26% from model 2) for β-TT detection was lower than the conventional threshold.
Regression-based machine learning models have the potential to estimate HbA2 using routine blood indices, offering a cost-effective screening/triage before specialized assays. Such models are particularly valuable in populations with prevalent iron deficiency anemia (serum ferritin <12 ng/ml), where fixed HbA2 cutoffs are unreliable. Larger studies incorporating peripheral smear image analysis are needed to improve the predictive accuracy of models before they can be deployed clinically.