AI-driven drug resistance profiling in tuberculosis patients: A transfer learning approach

Area of Science:

• Medical imaging analysis
• Genomic data interpretation
• Artificial intelligence in healthcare

Background:

• Tuberculosis (TB) remains a leading infectious cause of death globally.
• Rising rates of multi-drug resistant (MDR-TB) and extensively drug-resistant TB (XDR-TB) complicate treatment.
• Current diagnostic methods for drug resistance are time-consuming or inaccessible in resource-limited areas.

Purpose of the Study:

• To develop and evaluate an AI-powered framework for rapid and accurate profiling of TB drug resistance.
• To integrate diverse data sources including chest X-rays, genomic mutations, and clinical variables for enhanced diagnostic performance.

Main Methods:

• Utilized pretrained Convolutional Neural Networks (CNNs) like ResNet50, DenseNet121, and EfficientNet for chest X-ray (CXR) feature extraction.
• Employed transformer-based encoders to model genomic mutations and dense layers for clinical variable encoding.
• Fused modality-specific embeddings using an attention-based architecture.

Main Results:

• The tri-modal model achieved 91.2% accuracy, 89.4% sensitivity, and 92.7% precision.
• Area Under the Curve (AUC) values exceeded 90% for isoniazid, rifampicin, and ethambutol resistance.
• Ablation studies demonstrated significant performance improvements from integrating multiple data modalities.

Conclusions:

• Transfer learning and multi-modal data integration accelerate and enhance the identification of TB drug resistance.
• The proposed framework offers a scalable, cost-effective, and adaptable solution for personalized TB treatment selection.
• Future work will focus on expanding datasets and conducting prospective studies for clinical validation.