Truace
Subscribe

CBC-based deep learning triage for thalassemia before genetic testing

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

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. 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…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-1959· cite & verify
ThaCoder: an explainable deep learning framework for thalassemia screening using routine complete blood count: a multi-center validation study
"NABL Accredited Pathology Lab and Diagnostic Centre in Kharghar, Navi Mumbai - Full Body Checkup, CBC, Thyroid Blood Test and Home Blood Collection" by Goleisureintl is licensed under CC BY 4.0. To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/.
Trace impact readingContested
P 68The P score combines the specificity and measured human impact of the grounded problem claim with the strength of this Trace’s cited sources.

Both sides are scored from claims and sources, not community votes.

G 65The G score combines the specificity and measured human impact of the grounded gain claim with the strength of this Trace’s cited sources.

In brief

Researchers built an interpretable deep learning tool called ThaCoder that uses routine complete blood count results to flag people who may have thalassemia. Nine blood parameters were condensed into three latent factors, which were combined with gender and fed to a neural network trained on 2317 records. The locked model was then tested on 1442 records from four independent centers.

The approach could help places where genetic testing is costly or hard to reach by directing confirmatory tests to higher-risk patients. Its weaker detection of silent α-thalassemia carriers means it is a triage aid and not a diagnosis. Absolute risk estimates need recalibration at each site, and the cost estimate is only a preliminary illustration.

Main points

  1. Internal testing gave an AUROC of 0.9598 and specificity of 0.9337.
  2. External AUROC ranged from 0.8771 to 0.9439 across four centers, with overall sensitivity of 0.8387.
  3. Sensitivity was 0.5270 (39/74) for silent α-thalassemia carriers versus 0.8692 (226/260) for non-silent carriers.
  4. SHAP analysis showed that reduced MCV and MCH with a raised red-cell count drove positive predictions, consistent with the microcytic, hypochromic phenotype.
  5. A simulation assuming 10% prevalence suggested referrals and costs could fall by about 76%, but it is not a formal economic evaluation.

The gain

A deep neural network using routine blood count values separated likely thalassemia cases from others in four external centers, and a simulation suggests it could cut genetic-testing referrals by about 76%.

The problem

The same model detected only about half of silent α-thalassemia carriers, so milder carriers are missed at the chosen threshold.

The rundown

The model uses a two-stage design. Exploratory factor analysis with Varimax rotation reduced nine continuous blood parameters to three factors, which lowered multicollinearity among the erythrocyte indices. These factors and gender formed a four-dimensional input to a deep neural network tuned with a Bayesian-Grid search and out-of-fold threshold selection.

Validation was designed to avoid optimistic estimates. The model and operating threshold were locked, then evaluated without refitting or threshold reselection at four external centers, and between-center heterogeneity was quantified. Performance varied by center, with AUROC from 0.8771 to 0.9439.

The authors report that the SHAP-based explanations match the known blood profile of thalassemia, which supports interpretability. They also say the cost-reduction figure comes from a simulation at an assumed 10% prevalence, and that the model's absolute risk estimates need site-specific recalibration.

What this doesn’t fix

Sensitivity for silent α-thalassemia carriers was only 0.5270, the cost estimate comes from a preliminary simulation, and absolute risk estimates need site-specific recalibration.

Sources

  1. Peer-reviewedClinica Chimica Acta2026-10-17

The debate