Artificial intelligence-guided ultrasound lacks sufficient accuracy for deep vein thrombosis detection
1. Curry and colleagues tested an artificial intelligence (AI) software that guided non-radiology specialists in diagnosing proximal deep vein thrombosis (DVT).
2. The AI-guided ultrasound demonstrated failed to demonstrate safe accuracy for diagnosing proximal DVTs, and further software optimization is needed.
Evidence Rating Level: 2 (Good)
Study Rundown: DVTs are a leading cause of death and disability. However, an accurate diagnosis via compression ultrasound (CUS) often relies on trained, but scarce radiologists. AI-guided diagnostic tools have the potential to deliver care more efficiently in a resource-limited setting. Curry and colleagues conducted a multi-center, double-blind, test accuracy study on AutoDVT, an AI software that guided non-radiology specialists to perform CUS and diagnose DVTs. Adult patients with symptoms suggestive of DVTs, and if the diagnostic algorithm indicated the need for a CUS underwent an AutoDVT scan. The primary outcome was AutoDVT’s sensitivity compared to the one conducted by a radiologist as a reference standard. Secondary outcomes included the specificity, positive, and negative predictive values (PPV and NPV) of AutoDVT. The study defined the sensitivity target as 90% and the specificity target as 60%. The study found that AutoDVT achieved a sensitivity of 68% and a specificity of 80% for detecting proximal DVTs. For secondary outcomes, AutoDVT’s NPV and PPV were 95% and 28%, respectively. Overall, this study demonstrated that AutoDVT lacked sufficient test accuracy to be used safely in clinical practice compared to the reference standard.
Click here to read the study in NEJM AI
Relevant Reading: Artificial intelligence in thrombosis: transformative potential and emerging challenges
In-Depth [prospective cohort]: Curry and colleagues conducted a prospective, single-arm, double-blinded pilot study at eleven UK hospitals. Eligible patients were suspected of having DVTs and if the diagnostic algorithm indicated CUS. Patients who were pregnant beyond 12 weeks, not eligible for D-dimer testing, or had a previously confirmed DVT in the symptomatic leg, were excluded from the study. All participants had a Wells score calculated and a D-dimer test. Patients underwent an AutoDVT scan before a clinical CUS, and the participants, research staff, and remote reviewers were blinded to the result of the AutoDVT scan. The primary outcome was AutoDVT’s sensitivity compared to a reference standard of 90%. Secondary outcomes included the specificity, which was compared to a target of 60%; and PPV and NPV. In the end, 294 patients were analyzed, and AutoDVT achieved a sensitivity of 68% (95% confidence interval (CI), 49-83%) and a specificity of 80% (95% CI, 74-85%). However, the sensitivity increased to 85% (95% CI, 65-96%) when diagnosis was made using remote image analysis by clinicians. For NPV and PPV, the results were 95% (95% CI, 92-98%) and 28% (95% CI, 19-40%), respectively. Additionally, 81% of the initial AutoDVT scans had to be repeated for a median of three times due to failed scanning procedures. The study’s reliability benefited from its multi-center nature and the double-blind design, but inter-operator agreement analysis was not completed due to staff shortages. The authors concluded that the sensitivity of AutoDVT alone was too low for clinical use, and better-quality images are crucial for further software development.
Image: PD
©2025 2 Minute Medicine, Inc. All rights reserved. No works may be reproduced without expressed written consent from 2 Minute Medicine, Inc. Inquire about licensing here. No article should be construed as medical advice and is not intended as such by the authors or by 2 Minute Medicine, Inc.
