Machine learning for thalassemia detection: a critical review of diagnostic pathways and emerging technologies

Abstract

Thalassemia is still challenging to diagnose due to heterogeneous genotypes, overlapping hematologic phenotypes (e.g., other anemias), and limited access to standardized molecular testing. Current methods of diagnosis, like complete blood count (CBC) indices, smear analysis, electrophoresis/high-performance liquid chromatography (HPLC), and targeted genetics, work effectively in the clinic. However, they can be expensive and time-consuming when silent carriers or atypical variants are missing. This review shows the diagnostic pathway and, more importantly, highlights the necessity of emerging methods due to the limitations of current approaches. The focus in this study is on machine learning (ML) that learns from CBC features, analyzes smear and electrophoresis images, and integrates genetic and laboratory data. A summary of where ML helps (primary evaluation, inadequate confirmatory tests, better access in scarce resource settings) and its limitations are outlined (inadequate or biased datasets, weak external validation, poor calibration, unclear thresholds). It is suggested that the development of shared multicenter datasets, calibrated machine learning models with uncertainty estimates, and prospective studies reporting both clinical and economic impacts are essential. Rather than making detection the primary goal, the objective is to establish a balanced, data-driven diagnostic process that facilitates timely and individualized treatment. Future research should center on creating unified, data-driven approaches. These approaches should connect initial screening, confirmatory testing, and treatment planning within a single, cohesive system. The final goal is to create a fair learning health system. This system would continuously improve and make thalassemia risk evaluation, confirmation, and continual care available to diverse populations.