Transformer-based ensemble framework for tuberculosis treatment response prediction: multi-omics integration with clinical-grade performance

<h4>Background</h4>Tuberculosis (TB) treatment response prediction is critical for personalized care. We developed an ensemble Transformer framework integrating clinical and transcriptomic data from 467 patients and five GEO datasets (GSE83456, GSE107995, GSE158802, GSE19435, GSE25534).<h4>Methods</h4>Five diverse Transformer architectures were trained with Focal Loss, label smoothing, and stratified 5-fold cross-validation. Outcome labels were harmonized using sputum culture conversion at 2 months.<h4>Results</h4>The model achieved 97.1% accuracy and AUC 0.949 [95% CI: 0.912-0.978] on independent test folds, with sensitivity and specificity both 95%. Bootstrap validation (n = 1000) confirmed robustness.<h4>Conclusion</h4>This framework provides a clinically-relevant tool for TB treatment response prediction, pending prospective validation.<h4>Clinical trial registration</h4>https://www.chictr.org.cn/, identifier ChiCTR2300074328 03/08/2023.