Multicenter Double-Blind Study Evaluating AI-Driven Detection of Proximal Deep Vein Thrombosis

Background: Ultrasound is one of the most widely requested forms of diagnostic imaging. The costs for diagnosing deep vein thrombosis (DVT) in the UK are £175 million, annually. In at least 80% of cases, DVT is excluded. As health care provision becomes increasingly stretched, resource utilization needs to be optimized. This prospective, double-blind, test accuracy study was designed to test whether an artificial intelligence (AI)–guided software device (AutoDVT) could support nonradiology specialists to diagnose proximal DVT. Methods: Eleven regional hospital DVT diagnostic clinics enrolled adult patients, 18 years of age or older, who were referred for investigation of symptoms suggestive of DVT, including a compression ultrasound. Prior to the clinical compression ultrasound, an AutoDVT scan was completed. This was a two-point AI-guided compression ultrasound scan. We found that the main primary outcome was the sensitivity of AutoDVT within a diagnostic algorithm for the detection of proximal DVT by nonradiology-trained staff. Other outcomes included specificity and positive/negative predictive value of AutoDVT. Results: A total of 414 participants were enrolled. Proximal DVT was detected in 10.5% of those analyzed. AutoDVT resulted in 68% sensitivity (95% confidence interval [CI], 49 to 83%) and 80% specificity (95% CI, 74 to 85%) for the detection of proximal DVT. The negative predictive value for AutoDVT was 95% (95% CI, 92 to 98%), with a positive predictive value of 28% (95% CI, 19 to 40%). Overall, 63 out of 294 results (21%; 95% CI, 17 to 27%) were discrepant compared with compression ultrasound. Conclusions: Though AI-guided ultrasound use can detect proximal DVT, test accuracy was not sufficient for this device to be used safely. Further optimization of the software is required prior to use in clinical practice by nonradiology-trained health care professionals. (Funded by the Wellcome Trust [Wellcome Innovator Award 220505/Z/20/Z]. The trial was registered as ISRCTN 11069056.)