Ai models show varied accuracy for detecting pulmonary embolism in real hospitals
Multi-Site Real-World Performance of Commercial AI for Pulmonary and Incidental Pulmonary Embolism Detection
Computer Vision and Pattern RecognitionArtificial Intelligence
Summary
Pulmonary embolism is a serious condition causing blood clots in the lungs, but detecting it accurately can be hard. The authors looked at two AI programs used in hospitals to find these clots in lung scans. They found the AI was very good at spotting large, recent clots but often missed smaller or older ones. This means current AI tools work well in some cases but still need improvement and careful monitoring when used in real-world healthcare settings.
What this means in practice
- •For hospital radiology teams: Use the evaluated AI algorithms to assist in prioritizing suspected pulmonary embolism cases during CT scans, improving workflow efficiency with awareness of sensitivity limits.
- •For medical device regulators: Design post-market surveillance protocols for AI-based medical devices to track performance variations and safety in real-world clinical environments.
Authors
Aawez Mansuri, Mohammadreza Chavoshi, Theodorus Dapamede, Wasif Bala, Beatrice Brown-Mulry, Rohan Isaac, Bardia Khosravi, Hanzhou Li, Frank Li, John T. Moon, Chad Robichaux, Dan I. G. Cohen-Addad, Ninad V. Salastekar, Janice Newsome, Judy W. Gichoya, Hari Trivedi
Abstract
Pulmonary embolism (PE) is a leading cause of cardiovascular mortality, yet the real-world performance of FDA-cleared AI detection models remains incompletely characterized. We retrospectively evaluated two FDA-cleared AI algorithms from a single commercial platform (Aidoc Medical BriefCase), one for PE triage on dedicated CT pulmonary angiography (CTPA; n = 30,678) and one for incidental PE (iPE) detection on routine contrast-enhanced CTs (n = 37,191), across a 17-facility academic health system. Reference-standard labels were extracted from radiology reports using a validated LLM pipeline (97% accuracy, kappa = 0.94). The PE model achieved 86.8% sensitivity and 99.1% specificity, with sensitivity declining from 99.3% for saddle emboli to 72.9% for subsegmental PE, and from 89.7% for acute to 65.3% for non-acute PE. The iPE model achieved 73.5% sensitivity and 99.8% specificity. Both models demonstrated lower sensitivity than FDA-clearance benchmarks while exceeding cleared specificity, with diminishing performance for peripheral and non-acute emboli mirroring known human reader limitations and underscoring the need for standardized post-market surveillance of AI-enabled medical devices.