# P-1390. Accuracy of computer-aided detection of chest x-rays for pulmonary tuberculosis among adults with Xpert Ultra trace-positive sputum

## TL;DR

This study evaluates how well AI-based chest X-ray analysis helps diagnose tuberculosis in patients with weakly positive sputum tests, finding it has moderate accuracy but limitations.

## Contribution

The study provides new evidence on the diagnostic accuracy of AI-based chest X-ray analysis for TB in patients with trace-positive sputum results.

## Key findings

- CAD had an AUC of 0.65 for diagnosing TB using both composite and bacteriologically confirmed definitions.
- An X-ray score threshold of 0.5 had 66% sensitivity and 54% specificity for bacteriologically confirmed TB.
- CAD performance was higher among patients without prior TB treatment or HIV infection.

## Abstract

The clinical significance of low-level positive results from molecular testing of sputum for tuberculosis (TB) remains uncertain, and additional diagnostic testing, such as chest X-ray, might help to guide patient care. In high-TB-burden settings, where radiologists are often unavailable, artificial intelligence-based computer-aided detection (CAD) may be used to analyze chest X-rays, but its accuracy in this context remains unclear.Table 1.TB status determination at three months among outpatients with Xpert Ultra trace-positive sputum – overall and by X-ray score category and/or clinical risk subgroup, using two definitions of TB based on treatment decision with or without bacteriological confirmation.Figure 1.Receiver operating characteristic curves showing the performance of computer-aided detection software (qXR v4) on a baseline X-ray, for classifying bacteriologically confirmed and/or unconfirmed tuberculosis among individuals with an initial trace-positive sputum Xpert Ultra diagnostic result and up to three months of further diagnostic evaluation.

TB status determination at three months among outpatients with Xpert Ultra trace-positive sputum – overall and by X-ray score category and/or clinical risk subgroup, using two definitions of TB based on treatment decision with or without bacteriological confirmation.

Receiver operating characteristic curves showing the performance of computer-aided detection software (qXR v4) on a baseline X-ray, for classifying bacteriologically confirmed and/or unconfirmed tuberculosis among individuals with an initial trace-positive sputum Xpert Ultra diagnostic result and up to three months of further diagnostic evaluation.

We evaluated CAD accuracy among adults who presented to clinics in Uganda and South Africa and tested trace-positive on sputum Xpert Ultra (“Xpert”). Participants underwent repeat sputum Xpert, two mycobacterial cultures, chest X-ray, and HIV testing, and those not started on treatment after this baseline evaluation repeated sputum Xpert and culture testing at 1 and 3 months to further clarify TB status. Chest X-ray images were retrospectively analyzed by CAD software (qXR v4, Qure.ai). We evaluated CAD accuracy compared to two TB definitions: a composite definition that included clinical diagnoses, and a second requiring bacteriological confirmation.Figure 2.Distribution of X-ray scores from computer-aided detection software (qXR v4), among individuals with trace-positive sputum at clinics in Uganda and South Africa, colored by TB diagnostic outcomes and stratified by history of previous tuberculosis treatment and by HIV status.

Distribution of X-ray scores from computer-aided detection software (qXR v4), among individuals with trace-positive sputum at clinics in Uganda and South Africa, colored by TB diagnostic outcomes and stratified by history of previous tuberculosis treatment and by HIV status.

Among 276 participants, median age was 37 years (interquartile range 30-46), 141 (51%) were female, 156 (57%) were HIV-positive, and 102 (37%) had prior treatment for TB. The median X-ray score was 0.63 (interquartile range 0.14-0.91), and 153 (55%) had a score above 0.5 (the manufacturer-recommended threshold). TB was diagnosed in 131 patients (47%), including 76 with bacteriological confirmation. CAD had an area under the curve (AUC) of 0.65 (95% confidence interval 0.58–0.71) using a composite definition and 0.65 (0.57–0.73) for bacteriologically confirmed TB. An X-ray score threshold of 0.5 had sensitivity and specificity of 66% and 54%, respectively, for bacteriologically confirmed TB. Lowering the threshold to 0.2 increased sensitivity to 80% but decreased specificity to 37%. The AUC of CAD was higher among patients without prior TB treatment and among those without HIV infection (Figure 1).

Chest x-ray with CAD could serve as a supplementary tool for diagnosing TB in adults with Xpert trace-positive sputum, particularly those without prior tuberculosis or HIV infection. However, a low X-ray score did not reliably rule out TB disease.

Paul K. Drain, MD, MPH, Abbott Diagnostics: Advisor/Consultant|Abbott Diagnostics: Grant/Research Support|Giner: Advisor/Consultant|OraSure: Advisor/Consultant|Revvity: Advisor/Consultant|Revvity: Grant/Research Support|Roche: Advisor/Consultant|Roche: Grant/Research Support