Evaluating the Benefits of Machine Learning for Diagnosing Deep Vein Thrombosis Compared to Gold Standard Ultrasound- A Feasibility Study

Objective:
The use of handheld ultrasound probes for point-of-care ultrasound (POCUS) is increasingly adopted in the assessment of proximal leg deep venous thrombosis (DVT). ThinkSono Guidance (previously referred to as 'AutoDVT'), an artificial intelligence (AI)-app, has the potential to enable non-specialists to perform POCUS DVT scans and capture ultrasound sequences.
Once uploaded onto a cloud-dashboard, these sequences could enable remote diagnosis by a specialist. The study's objective is to evaluate the feasibility of remote diagnosis using ultrasound sequence acquisition and to assess the app's potential as a risk triaging tool.

Method:
Patients with suspected DVT were enrolled in a German hospital over a period of 3,5 month. Each participant underwent first an AI-guided 2-region POCUS conducted by a healthcare provider (HCP) without formal ultrasound training using a handheld ultrasound probe connected to the AI-app, before undergoing a formal DVT diagnostic scan. The app-generated sequences were anonymously reviewed by 5 remote qualified specialists on a cloud-dashboard. Reviewers evaluated the quality of all sequences using the American College of Emergency Physicians (ACEP) image quality scale (scoring from 1 to 5, with a score of ≥ 3 indicating adequate imaging quality). Sequences with an ACEP score ≥ 3, were categorised as "compressible," "incompressible," or "indeterminate." Sensitivity and specificity were calculated for scans with an ACEP score ≥ 3 and categorised as "compressible" or "incompressible." The "indeterminate," "low quality," and "incompressible" scans were combined into a ‘high-risk’ group.

Result:
In the study, 91 participants were included, with 62% falling within the 70-95 age group, and 59% being females. Out of these, 18% of the scans were incomplete. Among the remaining 75 scans, 91% had sufficient image quality, with an average ACEP score of 3.54. 10 of these 68 scans were categorised as "indeterminate." This classification occurred because reviewers either couldn't establish a definitive diagnosis or identified other pathologies, such as a Baker Cyst. Overall, the remote clinician effectively categorised 64% as either "compressible" or "incompressible."

The diagnostic accuracy for these 58 scans revealed a sensitivity of 100% and specificity of 91% when compared to the formal scan. Specifically, 17% of scans were identified as non-compressible, 83% as compressible. Noteworthy is that overall, 53% of scans were classified as low risk, suggesting that scans of this category might not necessitate additional formal scans.

Conclusion:
The AI-app proficiently directed non-experts in capturing valid ultrasound images, streamlining the process of remote triaging. This innovative approach has the potential to expedite DVT diagnosis and treatment, thereby reducing the necessity for formal scans, especially when a significant portion of them results in negative findings. Furthermore, this technology could not only enhance the efficiency of diagnostic procedures but also extends diagnostic capabilities into primary care settings.