Development and Internal Validation of a Machine Learning-Based Model for Thalassemia Identification Using Complete...

Background/Objectives: While genetic testing for thalassemia is definitive yet costly for population screening, and conventional methods lack specificity, this study aims to develop and internally validate a machine learning (ML) model based on routine complete blood count (CBC) parameters to identify individuals at increased risk of thalassemia who may benefit from confirmatory genetic testing. Methods: Data from 1820 individuals were retrospectively collected and randomly divided using stratified sampling into training (70%) and internal validation (30%) cohorts. Feature selection was conducted within the training cohort, reducing 225 analyzer-derived parameters to 12 features. SMOTE was applied to the training portion of each five-fold cross-validation split. Seven ML models were evaluated in the validation cohort, with post hoc global feature attribution assessed using SHAP. Results: Seven ML models were developed using 12 selected features. All performed strongly (AUC: 0.885-0.918), with Random Forest (RF) achieving the highest AUC of 0.918 (95% CI: 0.900-0.940), accuracy of 85.7%, sensitivity of 89.7%, specificity of 83.0%, and F1 score of 0.837, with a PPV of 78.4% and NPV of 92.1%. SHAP analysis identified mean corpuscular volume (MCV), mean corpuscular hemoglobin (MCH), and red blood cell distribution width-standard deviation (RDW-SD) as the most influential features. Conclusions: The RF-based model showed good discrimination in validation, while SHAP analysis provided information on global feature contributions. The model may assist in identifying individuals who warrant confirmatory testing; however, prospective external validation is required before routine clinical implementation.