Computer-aided detection (CAD) programs examine chest radiographs to identify signs of tuberculosis (TB) and generate a numerical abnormality score. However, assessments of CAD performance in TB screening are limited by insufficient confirmatory testing data, especially for individuals with lower scores or no symptoms. Moreover, optimal score cutoffs for triggering additional testing may vary by population traits and patient characteristics. This study sought to measure CAD accuracy across all screened people and determine if tailoring thresholds by age and sex could enhance results.This cross-sectional diagnostic accuracy study involved TB screening of individuals aged 15 years and older in Uganda using portable chest x-rays interpreted with CAD software (qXR version 3.2). Screening was offered to participants not receiving active TB treatment, irrespective of symptoms. We analyzed data from all eligible participants screened at facility- and community-based locations between June 1, 2022, and March 31, 2024. Those with scores at or above 0.1 (on a 0–1 scale) were requested to submit sputum samples for Xpert MTB/RIF Ultra testing. Diagnostic accuracy metrics (sensitivity, specificity, and area under the curve [AUC]) for detecting Xpert-positive TB were calculated using a uniform threshold, under varying assumptions about TB prevalence in those with scores below 0.1. These were then compared to stratified approaches based on age, sex, or both. Of 54,840 individuals evaluated for eligibility, 52,835 underwent CAD screening. Median age was 38 years (IQR: 26–50), with 23,586 (44.6%) males and 29,249 (55.4%) females. A total of 8,949 (16.9%) had scores of 0.1 or higher. Among 7,219 participants with valid Xpert results, 382 (5.3%) tested positive, including 81 trace results. Assuming 0.1% TB prevalence among those with scores below 0.1, qXR achieved an estimated AUC of 0.92 (95% CI: 0.90–0.94) for Xpert-positive TB. Stratification by age and sex improved performance; for instance, at 96.1% specificity (95% CI: 95.9–96.3), sensitivity reached 75.0% (69.9–79.5) with a universal threshold (≥ 0.65) versus 76.9% (71.9–81.2) with age- and sex-adjusted thresholds (P = 0.046). These results indicate that CAD accuracy for TB screening across all participants—including asymptomatic individuals and those with normal-appearing X-rays—is higher than previously reported. Customizing score thresholds based on factors such as age and sex may further improve performance, supporting a more targeted and efficient screening strategy.
