Remote Expert DVT Triaging of Novice-User Compression Sonography with AI-Guidance.

Background: Compression ultrasonography of the leg is established for triaging proximal lower extremity deep vein thrombosis (DVT). AutoDVT, a machine-learning software, provides a tool for nonspecialists in acquiring compression sequences to be reviewed by an expert for patient triage. The purpose of this study was to test image acquisition and remote triaging in a clinical setting.

Methods: Patients with a suspected DVT were recruited at 2 centers in Germany and Greece. Enrolled patients underwent an artificial intelligence-guided two-point compression examination by a nonspecialist using a handheld ultrasound device prior to a standard scan. Images collected by the software were uploaded for blind review by 5 qualified physicians. All reviewers rated the quality of all sequences on the American College of Emergency Physicians (ACEP) image quality scale (score 1-5, ≥ 3 defined as adequate imaging quality) and for an ACEP score ≥3, chose "Compressible", "Incompressible", or "Other". Sensitivity and specificity were calculated for adequate quality scans with an assessment as "Compressible" or "Incompressible". We define this group as diagnostic quality. To simulate a triaging clinical algorithm, a post hoc analysis was performed merging the "incomplete", the "low quality", and the "Incompressible" into a high-risk group for proximal DVT.

Results: Seventy-three patients (average age 64.2 years, 44% females) were eligible for inclusion and scanned by 3 nonultrasound-qualified healthcare professionals. Three patients were excluded from further analysis due to incomplete scans. Sixty two of 70 (88.57%) of the completed scans were judged to be of adequate image quality with an average ACEP score of 3.35. Forty seven of 62 adequate AutoDVT scans were assessed as diagnostic quality, of which 8 were interpreted as positive for proximal DVT by the reviewers resulting in a sensitivity of 100% and specificity of 95.12%. When simulating a triaging algorithm, 34/73 (46.58%) of patients would be triaged as high risk and 8 would be confirmed as positive for proximal DVT (6 in the diagnostic and 2 in the low-quality cohort). Of 39/73 patients triaged as low risk, all were negative for proximal DVT in standard duplex; thus, this triaging algorithm could potentially save 53.42% of standard duplex scans.

Conclusions: Machine learning software was able to aid nonexperts in acquiring valid ultrasound images of venous compressions and allowed remote triaging. This strategy allows faster diagnosis and treatment of high-risk patients and can spare the need for multiple unnecessary duplex scans, the vast majority being negative.