Performance of Human and Computer-aided Evaluation of Digital Chest Radiography for Community-based Screening of Asymptomatic Tuberculosis

Abstract Background The World Health Organisation (WHO) recommends digital chest radiography (dCXR) with computer-aided detection (CAD) for tuberculosis (TB) screening of individuals >15 years of age. Methodology Adults (≥18 years) were enrolled (March 2021-December 2022) in South Africa into a community-based Screening Cohort (household contacts) and a facility-based Triage Cohort (symptomatic clinic attendees). Microbiologically-confirmed pulmonary TB required positive sputum culture and/or Xpert Ultra. Asymptomatic TB was diagnosed in participants without TB symptoms. dCXR were read by blinded human readers and qXR CAD (0.5 threshold; Qure.AI, India). Results dCXR from 1,353 participants (886 Screening Cohort; 467 Triage Cohort) were analysed. Microbiologically-confirmed TB occurred in 48 (5.4%) Screening Cohort [9 symptomatic (19%) and 39 asymptomatic (81%)]; and 116 (24.8%) Triage Cohort (all symptomatic) participants. dCXR sensitivity (human readers) for asymptomatic TB in the Screening Cohort was 56.4%, vs. 72.4% for symptomatic TB in the Triage Cohort (difference -16%; 95%CI -2.9 to -29.1); with specificities 94.1% and 81.2%, respectively. Corresponding qXR CAD sensitivities were 69.2% vs. 83.6% (difference -14.4%; 95%CI -26 to –2.8), with specificities 89.3% and 73.5%, respectively. The difference in dCXR sensitivity and specificity for asymptomatic TB between qXR CAD and human readers was 12.8% (95%CI -0.48 to 26.1) and -4.8% (95%CI -12.4 to 28.2), respectively. Conclusion Sensitivity of community-based dCXR screening for microbiologically-confirmed asymptomatic TB among household contacts was lower than for facility-based triage of symptomatic TB, but approached 70% with CAD. Neither human reader nor qXR CAD evaluation met WHO targets for a TB screening test (90% sensitivity; 80% specificity). Research in context Evidence before this study The World Health Organisation (WHO) recommends digital chest radiography (dCXR) with computer-aided detection (CAD) for tuberculosis (TB) screening of individuals >15 years of age, based on data from prevalence surveys and facility-based studies. Performance data for community-based screening of asymptomatic TB are lacking. We searched PubMed for literature published in English between January 1, 2000, and November 1, 2025, for community-based, active case-finding studies of adolescents and adults aged 15 years and older that used dCXR CAD for asymptomatic TB screening. We used the following search terms: “Tuberculosis” AND (“asymptomatic” OR “subclinical”) AND (“computer aided diagnosis” OR “artificial intelligence”) AND “community-based screening” AND “chest radiography” AND (“diagnostic performance” OR “sensitivity”). We identified five studies reporting on microbiologically-confirmed asymptomatic TB and dCXR CAD performance. Three of five studies tested sputum only in those who were symptomatic and/or had abnormal CXR. One study did measure prevalence of asymptomatic TB by universal sputum testing of all participants, but did not report sensitivity and specificity for asymptomatic TB separately. One case-control study of CAD4TB (v7), which pooled data from five active case-finding cohorts, reported sensitivity of 61.4% and specificity of 86.7% for asymptomatic TB. However, the case-control design and inclusion of two cohorts using prevalence survey methodology and three cohorts enrolling high TB risk groups, two of which did not perform CXR on all participants, suggest potential for selection bias. Added value of this study We evaluated discriminatory performance of dCXR screening for asymptomatic TB among adult household contacts of TB patients, using human readers and qXR CAD (QURE.AI, India), in three communities in South Africa (Screening Cohort). Performance was benchmarked against that for symptomatic TB among adult clinic attendees (Triage Cohort), to enable comparison with traditional published approaches. All participants underwent universal sputum testing, regardless of symptom status or dCXR results. Sensitivity of human readers for asymptomatic TB in the Screening Cohort was 56.4%, compared to 72.4% for symptomatic TB in the Triage Cohort, with specificity 94.1% and 81.2%, respectively. The corresponding sensitivity of qXR CAD for asymptomatic TB, using the manufacturer’s 0.5 threshold score, was 69.2%, compared to 83.6% for symptomatic TB, with specificity 89.3% and 73.5%, respectively. The difference in dCXR sensitivity and specificity for asymptomatic TB between qXR CAD and human readers was 12.8% and -4.8%, respectively. The adjusted qXR threshold score (0.007) required to achieve 90% sensitivity for asymptomatic TB reduced specificity to 18.9%; and did not meet the WHO Target Product Profile (TPP) for a high sensitivity (90%), high specificity (80%) TB screening test. Implications of all the available evidence Sensitivity of community-based dCXR screening of household contacts for asymptomatic TB was low, compared to facility-based triage of symptomatic TB. Neither human reader nor qXR CAD evaluation of dCXR met the minimal WHO TPP for a high sensitivity (90%), high specificity (80%) TB screening test. Although dCXR CAD community screening would detect more than two-thirds of all people with previously undiagnosed, microbiologically-confirmed asymptomatic TB, the significant proportion of people with TB that would remain undetected, and untreated, might allow ongoing Mycobacterium tuberculosis transmission and hinder elimination efforts.