Beyond overall accuracy: subgroup performance and threshold calibration of AI-assisted chest radiography for equity-oriented pulmonary tuberculosis screening.
- 1Department of Intensive Care Unit, The First Affiliated Hospital of Zhejiang Chinese Medical University (Zhejiang Provincial Hospital of Chinese Medicine), Hangzhou, China.
Abstract
Artificial intelligence-assisted computer-aided detection (CAD) can expand access to chest radiography interpretation for pulmonary tuberculosis screening and triage, but high overall discrimination does not ensure safety at a deployed threshold. This focused narrative Mini Review synthesizes evidence on subgroup performance, threshold calibration, external validity, and implementation of CAD for adult pulmonary tuberculosis. Smear-negative, asymptomatic, and some HIV-associated disease may be less radiographically visible, increasing false-negative risk. Conversely, previous tuberculosis and competing chronic abnormalities reduce specificity and concentrate unnecessary confirmatory testing. Sex, age, geography, clinical setting, verification design, and software version can further alter operating performance. A numerical threshold is therefore not a portable algorithm property but a context-specific clinical resource-allocation policy. We propose a lifecycle in which programs define the intended use, lock the product and version, conduct representative local validation, select an operating point under explicit sensitivity and capacity constraints, audit clinically relevant subgroups, and monitor the complete care pathway. Stratified thresholds and multivariable models may improve selected settings, but current evidence does not justify replacing robust local calibration and governance. Prospective evaluations should compare threshold strategies under common verification standards and report missed cases, referral burden, treatment initiation, and patient-important outcomes.
Related
- Diagnostic Performance of Different AI Tools for Chest Radiography: A CT-Anchored Per-Abnormality Analysis in a Tertiary Referral Center.
- Diagnostic performance with sensitivity/specificity optimized computer-aided detection to detect common abnormalities in chest radiography
- Hydropneumothorax as a Diagnostic Clue to Pulmonary Tuberculosis With Pleural Involvement: A Case Report
- Determinants of Anticipated Video-Supported Treatment Acceptability in Drug-Susceptible Pulmonary Tuberculosis
- Pulmonary Tuberculosis and a Related Verrucous Plaque on the Sole of the Foot
BibTeX
@article{chen2026beyond,
title = {{Beyond overall accuracy: subgroup performance and threshold calibration of AI-assisted chest radiography for equity-oriented pulmonary tuberculosis screening.}},
author = {Wei Chen and Sixu Pan and Zhirong Zhang and Ying-Ying Shen and Zhaokun Fan and Cheng-En Li},
journal = {Frontiers in medicine},
year = {2026},
month = oct,
doi = {10.3389/fmed.2026.1980048},
url = {https://europepmc.org/article/MED/42953548},
note = {CC BY},
}
RIS
TY - JOUR
TI - Beyond overall accuracy: subgroup performance and threshold calibration of AI-assisted chest radiography for equity-oriented pulmonary tuberculosis screening.
AU - Wei Chen
AU - Sixu Pan
AU - Zhirong Zhang
AU - Ying-Ying Shen
AU - Zhaokun Fan
AU - Cheng-En Li
PY - 2026
DA - 2026/10/14
JO - Frontiers in medicine
DO - 10.3389/fmed.2026.1980048
UR - https://europepmc.org/article/MED/42953548
AB - Artificial intelligence-assisted computer-aided detection (CAD) can expand access to chest radiography interpretation for pulmonary tuberculosis screening and triage, but high overall discrimination does not ensure safety at a deployed threshold. This focused narrative Mini Review synthesizes evidence on subgroup performance, threshold calibration, external validity, and implementation of CAD for adult pulmonary tuberculosis. Smear-negative, asymptomatic, and some HIV-associated disease may be less radiographically visible, increasing false-negative risk. Conversely, previous tuberculosis and competing chronic abnormalities reduce specificity and concentrate unnecessary confirmatory testing. Sex, age, geography, clinical setting, verification design, and software version can further alter operating performance. A numerical threshold is therefore not a portable algorithm property but a context-specific clinical resource-allocation policy. We propose a lifecycle in which programs define the intended use, lock the product and version, conduct representative local validation, select an operating point under explicit sensitivity and capacity constraints, audit clinically relevant subgroups, and monitor the complete care pathway. Stratified thresholds and multivariable models may improve selected settings, but current evidence does not justify replacing robust local calibration and governance. Prospective evaluations should compare threshold strategies under common verification standards and report missed cases, referral burden, treatment initiation, and patient-important outcomes.
SN - 2296-858X
ER -