Recent advances in artificial intelligence across interventional pulmonology: a narrative review

Background and Objective: Artificial intelligence (AI) is increasingly applied across interventional pulmonology (IP), though the transition from technical innovation to actual clinical benefit remains a work in progress. This narrative review examines the current evidence of key AI subtypes including machine learning (ML), deep learning (DL), and radiomics, and their roles across the major domains of IP such as lung cancer screening, bronchoscopy, endobronchial ultrasound (EBUS), pleural disease, and therapeutic airway interventions. Methods: A structured search was conducted across PubMed/MEDLINE, Embase, Web of Science, Scopus, ScienceDirect, and Google Scholar for studies published between April 2006 and January 2026 using terms related to AI and IP. Preclinical, retrospective, prospective, and translational studies were eligible. Given the heterogeneity in study designs, AI methodologies, and outcome measures, findings were synthesized qualitatively. Key Content and Findings: In lung cancer screening, AI applied to low-dose computed tomography (LDCT) and chest radiography (CXR) improves nodule detection and malignancy risk stratification. Radiomics extends these capabilities by characterizing lesion biology, predicting molecular features, guiding biopsy targeting, and improving mediastinal staging. Multimodal approaches using radio-pathomic signatures and liquid biopsy data have also been shown to improve prognostication and personalized decision-making. In bronchoscopy, AI has been studied in airway image interpretation, lesion classification, real-time navigation, and trainee assessment. EBUS-based AI models improve lymph node and peripheral lesion characterization, while AI-assisted rapid on-site evaluation (ROSE) approaches expert-level performance in specimen adequacy assessment, malignant cell detection, and cytologic subtyping. In pleural disease, AI tools combining imaging, ultrasound (US), cytology, and clinical data further improve the detection and classification of effusions. Therapeutically, AI-driven quantitative imaging supports patient selection and procedural planning in bronchoscopic lung volume reduction (BLVR), airway stenosis and stenting, and prediction of tracheostomy timing and decannulation readiness. Despite these advancements, most AI applications remain investigational, with evidence largely derived from retrospective, single-center studies and limited impact on meaningful clinical outcomes. Conclusions: AI holds genuine potential to enhance diagnosis and procedural guidance across IP. However, its clinical utility in routine practice still requires prospective multicenter validation, standardized datasets, workflow integration, cost-effectiveness assessment, explainability, and human-centered implementation. AI should currently be viewed as an adjunct to, rather than a replacement for, interventional pulmonologists.