Combining Radiological and Genomic TB Portals Data for Drug Resistance Analysis

Area of Science:

• Medical Informatics
• Genomics
• Radiology

Background:

• Drug-resistant tuberculosis (DR-TB) poses a growing global health threat, with treatment success rates around 60%.
• Early detection of drug resistance is critical for patient outcomes and controlling transmission.
• The TB Portals program offers a valuable, albeit heterogeneous, dataset of DR-TB cases, integrating clinical, genomic, and radiological information.

Purpose of the Study:

• To evaluate the utility of combining host radiological features (chest X-rays) and pathogen genomic features for identifying drug susceptibility and predicting treatment duration in tuberculosis.
• To address challenges posed by imbalanced and high-dimensional data within the TB Portals dataset.

Main Methods:

• Utilized multi-modal data from the TB Portals program, including chest X-ray derived radiological features and pathogen genomic data.
• Developed and evaluated classification models for distinguishing drug-sensitive (DS-TB) and drug-resistant tuberculosis (DR-TB).
• Constructed regression models to predict the length of the first successful drug regimen, comparing performance using radiological, genomic, or combined features.

Main Results:

• A classification accuracy of 92.4% was achieved for DR-TB/DS-TB identification using genomic features alone or combined with radiological features.
• Combining radiological and genomic features reduced the relative error in predicting the length of the first successful treatment to 22.0%, compared to 25.6% for genomic features alone.
• Radiological features alone yielded the best performance (17.8% relative error) for predicting the treatment length of the most common drug combination.

Conclusions:

• While combining radiological and genomic data did not enhance DR-TB/DS-TB classification accuracy, it significantly improved the prediction of treatment length.
• Radiological features alone demonstrated superior performance in predicting treatment duration for specific common drug regimens.
• The study highlights the potential of integrating diverse data modalities for optimizing tuberculosis management and patient care.