An automated tuberculosis screening strategy combining X-ray-based computer-aided detection and clinical information

Area of Science:

• Medical Imaging
• Machine Learning
• Public Health

Background:

• Tuberculosis (TB) screening programs in endemic regions face challenges due to limited human resources and radiological expertise.
• Computer-aided detection (CAD) offers a potential solution for chest radiograph (CXR) interpretation.
• Existing automated methods do not integrate readily available clinical information.

Purpose of the Study:

• To develop and evaluate a machine learning-based framework that combines CAD scores from CXRs with clinical features for improved TB screening.
• To assess the performance of this integrated approach against standalone CAD and clinical information strategies.

Main Methods:

• A machine learning combination framework was developed to integrate CAD scores and 12 clinical features.
• The framework was evaluated on a prospective dataset of 392 suspected TB patients from Cape Town, South Africa.
• Performance was compared using area under the ROC curve, specificity at 95% sensitivity, and negative predictive value.

Main Results:

• The combination framework achieved a higher area under the ROC curve (0.84) compared to CAD alone (0.78) and clinical information alone (0.72).
• Specificity at 95% sensitivity was significantly improved (49%) with the combined approach versus individual methods (24% and 31%).
• The negative predictive value was also higher for the combination framework (98%) compared to standalone strategies (95% and 96%).

Conclusions:

• Combining CAD and clinical information presents a promising strategy for enhancing the accuracy of TB screening.
• This integrated approach can help mitigate challenges posed by resource limitations in TB-endemic areas.
• The developed framework offers a valuable tool for estimating the risk of active TB disease during screening.