ABSTRACT
Efthymios Avgerinos, MD
Co-Director Clinic of Vascular and Endovascular Surgery
Athens Medical Center - University of Athens
Marousi Athens, Attiki, Greece
Compression ultrasonography of the leg is established for triaging proximal lower extremity deep venous thrombosis (DVT) when performed by a certified operator. It can be associated with long waiting times and high resource utilization. AutoDVT, a machine-learning software, provides a tool for non-specialists in acquiring appropriate two or three point compression sequences. These can be reviewed by an expert to triage patients. 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 two tertiary centers. Enrolled patients underwent an AutoDVT scan by a non-ultrasound-qualified professional prior to the standard duplex scan. A handheld device with proprietary image-acquisition software to perform a two-point-compression examination (groin and popliteal fossa) was used. Images collected by the software were uploaded to a cloud-based platform for blind review by four external qualified physicians. All reviewers, based on these images, rated the overall quality of all sequences on the ACEP scale (score 1-5, ≥3 defined as sufficient diagnostic-quality) and if sufficient chose “DVT” vs “No DVT”. Sensitivity and specificity were calculated based on the “sufficient-quality” scans (score ≥3). To simulate a triaging clinical algorithm, an analysis was performed merging the “low quality”, the “incomplete” and the “incompressible vein” scans as high risk for proximal DVT. These were then compared to the standard duplex diagnosis.
Results:
73 patients (64.2±17.9 age, 27.7±6.2 BMI, 44% females) were eligible for inclusion and scanned by 3 nurses. Three patients were excluded from further analysis due to incomplete scans. 54 (77%) of the completed scans were of diagnostic quality with an average ACEP score of 3.40 ± 0.86, 12 of which were interpreted as positive for proximal DVT; 6 of them were confirmed with standard duplex and 6 were false positives (sensitivity 100%, specificity 87.50%). The remaining 42 scans were all negative for proximal DVT; 4 isolated calf DVTs were identified in standard duplex. When simulating a triaging algorithm, 31/73 (42%) patients would be triaged as high risk and 8 would be confirmed as positive for proximal DVT. Of 42 patients triaged as low risk, all were negative for proximal DVT in standard duplex, thus including AutoDVT in a proximal-DVT triaging algorithm could save 58% of standard duplex scans.
Conclusions:
Machine learning software was able to aid non-experts in acquiring valid ultrasound images of venous compressions and allowed safe and efficient remote triaging. Such a triaging strategy allows faster diagnosis and treatment of high-risk patients and can spare the need and cost of multiple unnecessary duplex scans.