Tuberculosis disease severity assessment using clinical variables and radiology enabled by artificial intelligence

Area of Science:

• Pulmonary Medicine
• Radiology
• Artificial Intelligence in Healthcare

Background:

• Current methods for assessing tuberculosis (TB) severity via chest X-rays (CXR) lack standardized radiological metrics for predicting treatment outcomes.
• Personalized TB treatment, including shortened regimens, requires accurate stratification of patient groups based on disease severity.
• Digital CXR and AI offer potential for improved TB diagnosis, monitoring, and individualized treatment approaches.

Purpose of the Study:

• To systematically associate baseline CXR metrics with TB treatment outcomes in diverse real-world clinical settings.
• To evaluate the utility of the percent of lung involved (PLI) as a radiological metric for predicting unfavorable TB treatment outcomes.
• To develop and validate an AI-based method for automated PLI measurement from CXR images.

Main Methods:

• Logistic regression analysis of 10 CXR metrics (including PLI, cavitation, Timika score) and clinical variables in 2,809 patients, stratified by drug resistance and HIV status.
• Fine-tuning of convolutional neural networks (CNNs) for automated PLI measurement from 5,261 CXR DICOM images.
• Comparison of predictive performance (AUC) of PLI against other radiological metrics and clinical models.

Main Results:

• PLI was the only CXR finding associated with unfavorable outcomes across all subgroups (e.g., adjusted OR 1.11 for drug-susceptible TB without HIV).
• PLI alone predicted unfavorable outcomes as effectively as or better than Timika score or cavitation (AUC PLI 0.656 vs. Timika 0.655, cavitation 0.591).
• A PLI threshold of >25% showed better outcome separation than >50%. The best CNN ensemble achieved an AUC of 0.850 for predicting PLI >25%.

Conclusions:

• Percent of lung involved (PLI) is a robust radiological marker for predicting unfavorable TB treatment outcomes, superior to cavitation and other metrics.
• Automated PLI measurement using CNNs is accurate and feasible for real-world application.
• Combining PLI with clinical factors like age, sex, and smear grade optimizes prediction of unfavorable pulmonary TB treatment outcomes.