Abstract / Summary
Artificial intelligence (AI) in pulmonology comprises distinct tools for thoracic imaging, tuberculosis screening, pulmonary function testing, risk estimation, and clinical information management. Maturity differs substantially by task and outcome. The strongest evidence concerns assisted interpretation of chest radiographs and computed tomography. Even in these domains, retrospective discrimination does not establish patient benefit or readiness for routine implementation. A randomized clinical trial in respiratory outpatient clinics showed that AI assistance improved non- radiologist physicians' detection performance on chest radiographs but no significant change in clinical decisions was observed. In Brazil, retrospective validation found internal AUCs of 0.94 for pulmonary abnormality and tuberculosis and performance comparable with physicians of heterogeneous experience during external assessment; clinical outcomes and routine deployment were not evaluated. A Brazilian lung cancer screening study reported 92.5% sensitivity and 97.8% negative predictive value for classifying scans as positive (Lung-RADS 3 or 4) against radiologist assessment, alongside 78.5% specificity, 50% positive predictive value, and missed nodules, reinforcing the need for radiologist oversight. Tuberculosis is a relevant exception because the World Health Organization endorses computer-aided detection software for digital chest radiography screening in people aged 15 years or older. A positive screen still requires confirmatory diagnostic evaluation, and thresholds require contextual calibration. WHO does not currently recommend these systems for screening children younger than 15 years. Brazilian pediatric study reported 52% sensitivity against microbiological confirmation, illustrating why adult indications must not be extrapolated. For COPD, asthma, interstitial lung disease, obstructive sleep apnea, bronchoscopy, and event prediction, development studies, retrospective validation, and heterogeneous reviews predominate. Evidence of prospective clinical utility, workflow and outcome impact, subgroup safety, and sustained implementation remains incomplete. Direct Latin American clinical evidence identified in this review was narrow and concentrated in Brazil. PAHO reports regional progress in connectivity, interoperability, governance, and public-sector discussion of AI, together with persistent infrastructure, data, workforce, and regulatory gaps. International results should therefore be treated as implementation hypotheses that require local validation. The responsible next step is a staged regional research agenda rather than a general recommendation for clinical pilots. Keywords: artificial intelligence; pulmonology; chest radiography; computed tomography; tuberculosis; pulmonary function; Latin America; clinical validation; governance.