ThaCoder: an explainable deep learning framework for thalassemia screening using routine complete blood count: a multi-center validation study.

BACKGROUND AND AIMS: Thalassemia remains a global public health challenge, yet definitive diagnosis via genetic testing is often restricted by high costs and infrastructure requirements. We aimed to develop and externally evaluate an interpretable deep-learning framework using routine complete blood count (CBC) parameters for thalassemia screening.

MATERIALS AND

METHODS

Using 2317 CBC records from the internal derivation cohort, we implemented a two-stage analytical architecture. First, exploratory factor analysis (EFA) with Varimax rotation reduced nine continuous parameters to three factors, providing a lower-dimensional representation that attenuated multicollinearity among the erythrocyte indices. These latent factors, alongside gender, formed a 4-dimensional feature space that was processed through a deep neural network (DNN) optimized via a two-stage Bayesian-Grid search strategy and out-of-fold threshold selection. The locked model and operating threshold were evaluated without refitting or threshold reselection across four independent external centers (N = 1442).

RESULTS

In the internal test set, the model achieved an AUROC of 0.9598 and a specificity of 0.9337. External AUROC ranged from 0.8771 to 0.9439 across the four centers. Sensitivity was 0.8387 overall but was lower for silent α-thalassemia carriers (0.5270, 39/74) than for non-silent α-thalassemia carriers (0.8692, 226/260). SHAP-based interpretability indicated that the model's learned associations are consistent with the microcytic, hypochromic phenotype of thalassemia, with reduced MCV and MCH alongside a raised red-cell count emerging as the dominant drivers of a positive prediction. Health economic simulations suggest this algorithmic triage could reduce genetic-testing referrals and associated costs by approximately 76% in a preliminary 10%-prevalence illustration, although a subset of milder carriers is missed at the chosen threshold and the estimate is not a formal economic evaluation.

CONCLUSIONS

By integrating statistical feature engineering with deep learning, this framework provides a CBC-based pre-genetic-testing risk-stratification (triage) approach. The locked model and operating threshold were evaluated without refitting in four external centers, with between-center heterogeneity quantified, although absolute risk estimates require site-specific recalibration.