COIOSDeep risk insights through multi-disciplinary research across the globe
cable network
Item new Preprint · not peer-reviewed Added

FDA-cleared AI triage for pulmonary embolism detected 64% of emboli in routine use

In 14,736 CT pulmonary angiograms, the Aidoc triage tool's sensitivity was 64.1%, against 93% at FDA clearance.

This retrospective study assessed the Aidoc BriefCase pulmonary embolism triage model on 14,736 CT pulmonary angiograms at six sites. Scans were from July 2023 to March 2026, with labels extracted from reports by a large language model. Sensitivity was 64.1% and specificity 98.8%, against 93% and 93.7% in the FDA 510(k) summary. Sensitivity was 33.3% for subsegmental emboli and differed between the two busiest sites, 68.7% versus 55.7%. Over time the false-positive rate fell 40% a year, while sensitivity at one site fell from 72.5% to 65.0%. Labels rest on report text, and the study cannot say whether this shift reflects data drift or threshold changes.

Why it is interesting: Deployed sensitivity fell well short of the cleared benchmark and differed between sites. False alarms fell and sensitivity drifted down, with no model or threshold change reported.

Source
medRxiv, 12 November 2026
Paper
Silent Drift and Site-level Variance in Clinical AI: Real-world performance and subgroup analysis for post-market monitoring of an FDA-cleared artificial intelligence (AI) triage tool for high-risk use case of pulmonary embolism
Authors
Dayma K, Newton C, Kawzowicz M, Patel P, Curl R, Ogden K
DOI
10.64898/2026.11.10.26367739
Type
Preprint
Design
Retrospective six-site health-system evaluation of 14,736 CT pulmonary angiograms, July 2023–March 2026, with LLM-extracted reference labels validated against 600 radiologist-annotated reports
Verdict
New finding
Driver
AI in medicine

Other drivers in Data, models and systems