Transformer-based ensemble framework for tuberculosis treatment response prediction: multi-omics integration with...

Area of Science:

• Genomics
• Machine Learning
• Infectious Disease Research

Background:

• Personalized medicine for tuberculosis (TB) requires accurate treatment response prediction.
• An ensemble Transformer framework was developed integrating clinical and transcriptomic data.
• Data from 467 patients and five Gene Expression Omnibus (GEO) datasets were utilized.

Purpose of the Study:

• To develop and validate a novel computational framework for predicting tuberculosis treatment response.
• To integrate diverse data types, including clinical and transcriptomic information, for improved prediction accuracy.
• To establish a clinically relevant tool for guiding personalized TB treatment strategies.

Main Methods:

• Five distinct Transformer architectures were trained using Focal Loss and label smoothing.
• Stratified 5-fold cross-validation was employed for model training and evaluation.
• Outcome labels were harmonized based on sputum culture conversion at two months post-treatment initiation.

Main Results:

• The model demonstrated high performance with 97.1% accuracy and an Area Under the Curve (AUC) of 0.949.
• Sensitivity and specificity were both reported at 95% on independent test folds.
• Bootstrap validation (n=1000) confirmed the robustness and reliability of the prediction model.

Conclusions:

• The developed ensemble Transformer framework offers a promising, clinically relevant tool for predicting TB treatment response.
• The model's high accuracy and robustness support its potential application in personalized TB care.
• Prospective validation is recommended to further confirm the framework's clinical utility.