Deployment of a commercial artificial intelligence tool was associated with increased reporting of incidental pulmonary embolism on routine contrast-enhanced computed tomography, particularly at the segmental level.
Researchers conducted a single-center, retrospective cross-sectional study comparing 2 4-month periods at a large academic tertiary care hospital: September 1 to December 31, 2022, before artificial intelligence (AI) implementation, and the corresponding period in 2023 after implementation. The analysis included 11,690 contrast-enhanced computed tomography (CT) examinations in consecutive patients aged 16 years or older, including 10,411 non–CT pulmonary angiography examinations with complete or partial chest coverage. Radiology reports were searched for terms indicating pulmonary embolism, and those identified were manually reviewed to confirm cases.
The primary analysis assessed the reporting rate of incidental pulmonary embolism (iPE) before and after AI deployment. Investigators also examined embolus location, CT protocol coverage, field of view, contrast phase, and AI performance. Pulmonary emboli identified on dedicated CT pulmonary angiography were classified as symptomatic, whereas those detected on other contrast-enhanced CT examinations performed for indications other than suspected pulmonary embolism were classified as incidental.
Among eligible non–CT pulmonary angiography examinations, the iPE reporting rate increased from 0.35% (17 of 4,888 examinations) before AI implementation to 0.63% (35 of 5,523) after implementation. The clearest change occurred at the segmental level, where iPE detection increased from 0.08% (4 of 4,888) to 0.42% (23 of 5,523). Detection of main, lobar, and subsegmental iPE did not change significantly.
The real-world evaluation also revealed incomplete AI coverage and modest sensitivity. The algorithm analyzed 4,344 of 5,523 eligible non–CT pulmonary angiography examinations during the postimplementation period, leaving 21.3% unanalyzed. Among 32 iPE cases in examinations processed by AI, the tool identified 21, corresponding to 65.6% sensitivity. When all eligible examinations, including those not processed by AI, were considered, real-world sensitivity was 60%.
Performance differed by contrast phase. AI sensitivity was 77.8% in late arterial-phase examinations, 71.4% in early arterial-phase examinations, and 56.2% in venous-phase examinations. CT coronary angiography and CT angiography of the aortic arch and carotids accounted for many of the unanalyzed examinations because of a local routing and series-selection issue that was corrected near the end of the study period.
The additional iPE findings frequently prompted changes inimmediate management. Overall, 77% (40 of 52) of patients with iPE had an immediate change in clinical management, primarily through the initiation or intensification of anticoagulation or placement of an inferior venacava filter. Most iPE cases occurred among inpatients, who accounted for 62.9% of cases after AI implementation, and 37% of patients with iPE had a known malignancy at the time of detection.
However, the researchers could not determine how much of the increased reporting resulted directly from AI detection vs heightened radiologist attention to the pulmonary arteries following AI integration. They also noted uncertainty about the clinical significance of additional, predominantly peripheral emboli detected with AI assistance.
Study limitations included the retrospective, single-center design, which may limit generalizability to institutions with different populations and imaging protocols. Reliance on radiology report keyword searches may have missed some cases; detailed comorbidities and venous thromboembolism risk factors were not systematically collected, and some CT angiographic examinations were not initially routed to the AI algorithm. The original radiology report also served as the reference standard for iPE diagnosis, and performance estimates stratified by contrast phase and chest coverage were based on small numbers of iPE cases.
Deployment of AI for pulmonary embolism detection may help identify additional incidental emboli during routine contrast-enhanced CT, particularly at the segmental level, while underscoring the importance of assessing coverage and performance under real-world conditions. The investigators called for prospective multicenter studies to determine the prognostic implications and appropriate management of additional iPE detected following AI implementation. “These findings suggest that AI may be uncovering a previously hidden burden of peripheral emboli,” wrote lead study author Fergus O’Herlihy, MB, BCh, BAO, of the Department of Radiology at Mater Misericordiae University Hospital in Dublin, Ireland, and colleagues.
The authors reported no study funding. Fionnuala Ni Ainle, PhD, disclosed payments from Boston Scientific for trial executive committee duties, paid to her institution, and investigator-initiated study payments from Boston Scientific, Actelion, Bayer, Sanofi, and Daiichi-Sankyo, paid to her university. The other authors reported no conflicts of interest.
Source: British Journal of Radiology
