Tuberculosis Disease Severity Assessment Using Clinical Variables and Radiology Enabled by Artificial Intelligence.

Area of Science:

• Radiology
• Infectious Disease Epidemiology
• Artificial Intelligence in Medicine

Background:

• Chest X-ray (CXR) is crucial for assessing pulmonary tuberculosis (TB) severity and guiding treatment duration.
• Optimal radiological metrics and their integration with clinical data for predicting TB treatment outcomes remain unclear.

Purpose of the Study:

• To evaluate radiological metrics from CXR for predicting unfavorable TB treatment outcomes.
• To determine the optimal combination of radiological and clinical variables for risk stratification.
• To develop an automated method for measuring a key radiological metric using artificial intelligence.

Main Methods:

• Logistic regression analysis of human-read and AI-generated CXR metrics in a real-world TB dataset (n=2809).
• Assessment of standalone predictive accuracy for 10 radiological features.
• Development and fine-tuning of convolutional neural networks (CNNs) to automate percent of lung involved (PLI) measurement from CXR images (n=5261).

Main Results:

• Human-read PLI was the only CXR finding associated with outcome across drug resistance and HIV subgroups.
• PLI demonstrated superior predictive accuracy compared to cavitation (AUC 0.654 vs 0.581) and outperformed commercial AI features.
• Combining PLI with age, sex, and smear grade improved outcome prediction (ΔAUC 0.028).
• A CNN ensemble achieved high accuracy (AUC 0.850) in predicting PLI >25%.

Conclusions:

• Percent of lung involved (PLI) is a superior radiological marker for predicting TB treatment outcomes compared to cavitation.
• Integrating PLI with clinical variables enhances risk stratification for TB patients.
• Automated PLI measurement using CNNs offers a scalable and accurate approach for clinical application.