Machine learning for biomarker-driven thalassemia diagnosis: a framework for causal effect estimation

Thalassemia is one of the most heterogeneous diseases, with more than a thousand mutation types recorded worldwide. Molecular diagnosis of thalassemia by conventional PCR-based DNA analysis is time- and resource-consuming owing to the phenotype variability, disease complexity, and molecular diagnostic test limitations. Thalassemia diagnosis relies heavily on complex interactions among hematological biomarkers, making it challenging to distinguish causal relationships from mere statistical correlations. To address this limitation, this study proposes a causal machine learning framework that integrates causal discovery and counterfactual inference to identify biomarker-driven treatment effects for thalassemia diagnosis. The proposed approach models clinical variables using a Directed Acyclic Graph (DAG) to capture underlying causal dependencies and enables robust estimation of treatment effects while accounting for confounding factors. To address the primary objective from complementary causal perspectives, the framework combines adjusted regression for confounding control, inverse probability weighting for propensity-based estimation, Random Forest-based analysis for nonlinear feature importance and heterogeneous treatment effects, and bootstrap resampling for uncertainty quantification. The adjusted linear regression model yielded a conservative causal effect estimate of 0.3008, while propensity-based methods demonstrated stronger effects, with IPW producing an estimate of 1.0604. The bootstrap approach confirmed the stability of these findings with a 95% confidence interval of [0.8806, 1.2161], indicating statistically significant treatment influence. Furthermore, heterogeneous effect analysis using CATE revealed substantial variability across individuals, with estimates ranging between 0.0848 and 2.5200, highlighting the importance of personalized assessment. Stratified analysis across age groups showed distinct differences, with the highest effect observed in adolescents (1.3484) and lower effects in children (0.8856), emphasizing demographic-specific variations. The results demonstrate that key biomarkers, particularly hemoglobin-related features, exert significant causal influence on thalassemia diagnosis.