Deployment-Oriented Evaluation of an Explainable Tabular Transformer for Tuberculosis Outcome Prediction Using Synthetic Low-Resource Healthcare Data

Résumé

Background and Purpose:  Healthcare systems use EHRs and predictive analytics in clinical decision-making processes; nevertheless, many current healthcare-based AI applications focus more on predictive accuracy than explainability, reliability, subgroup consistency, and deployment feasibility within limited healthcare settings. This study assesses an explainable transformer architecture in predicting TB treatment outcomes by employing a synthetic healthcare data set to conduct replicable experiments without compromising the privacy of patients’ information.

Methods: The proposed framework combines tabular machine learning with transformers, SHAP-based explainability, calibration tests, subgroup consistency tests, and evaluation for implementation into one AI pipeline in healthcare. Comparative experiments have been performed against Random Forest, XGBoost, and multilayer perceptron baselines on Accuracy, Precision, Recall, F1-score, and AUC-ROC measures. The transformer showed the highest performance in all the experiments, resulting in Accuracy 82.1%, F1-score 0.826, and AUC-ROC 0.872, as well as retaining decent calibration and subgroup consistency.

Results: Explainability analysis showed that HIV status, drug resistance, treatment adherence, nutrition, and healthcare access measures were significant contributors to the prediction performance, thus demonstrating that the model identified clinically meaningful interactions in the synthetic environment. These results should be viewed purely as an indication of technical feasibility, since all experiments took place in a controlled synthetic setting only.

Conclusions: Beyond predictive performance, the study contributes a deployment-oriented healthcare AI evaluation framework integrating explainability, calibration, subgroup inspection and operational feasibility assessment. This paper highlights the need for healthcare AI models that not only have predictive ability but are also trustworthy and feasible.