Artificial Intelligence-Supported Early Detection of Lung Cancer from Chest X-ray in Routine Clinical Practice: Real-World, Multicenter Study Across Czech Hospitals

Autoři

KVAKOVÁ Karolína KVAK Daniel DANDÁR Jakub ŠŤASTNÝ Marek OLEJKO Jakub KULTÁŇ Juraj GEROLD Jiří KOTOUČOVÁ Alena VOJTEK Cyril STRUNA Pavel DVOŘÁČKOVÁ Eva SMETANA Jiří František

Rok publikování 2025
Druh Konferenční abstrakty
Citace
Popis Background Although AI systems for chest X-ray (CXR) interpretation may enable earlier detection of thoracic malignancy in routine care, large-scale real-world evidence remains limited. We present results of a joint Carebot–Bristol Myers Squibb (BMS) program integrating AI decision-support into routine workflows to evaluate feasibility, flagged-case yield, and downstream actions for suspected thoracic malignancy (primary lung cancer or pulmonary metastases). Methods All CXRs across nine Czech hospitals (regional and tertiary; two pneumo-oncology centers) from Jan 1–Jun 30, 2025 were automatically analyzed with commercial-stage AI software (Carebot AI CXR; Carebot s.r.o., Czechia). The AI flagged abnormalities; a joint panel of hospital and interim Carebot radiologists reviewed and classified exams as suspicious or not for thoracic malignancy. Suspicious cases entered fast-track diagnostics. Primary outcomes: (i) proportion flagged as suspicious, (ii) patients referred for diagnostic work-up, and (iii) confirmed malignancies. Results Among 96 459 CXRs, the algorithm identified abnormalities in 16 030 exams (16.6%). Multidisciplinary review classified 837 CXRs (0.87%) as suspicious. Follow-up status was available in 561/837 (67.0%): 211 (25.2% of all suspicious; 37.6% of those with status) were referred for immediate diagnostic work-up, 350 (41.8% of all; 62.4% of recorded) were cleared without further work-up, and 276 (33.0%) remain under evaluation. Among the 211 investigated, 54 previously undiagnosed thoracic cancers were detected, 70 confirmed known malignancies or pulmonary metastases, 38 await confirmation of nodule origin (biopsy or PET/CT), 20 entered long-term radiologic surveillance, and 29 were recalled owing to missing clinical data. Conclusions In this single-country, real-world, multicenter analysis of routine CXR exams during H1-2025, the Carebot–BMS AI triage workflow enabled early identification of thoracic malignancies within routine care while triaging <1% of exams for expedited review. Ongoing follow-up and planned health-economic work will further characterize long-term value and scalability of this AI-supported workflow.

Používáte starou verzi internetového prohlížeče. Doporučujeme aktualizovat Váš prohlížeč na nejnovější verzi.