Keywords:
Artificial Intelligence, eHealth, Vascular, Ultrasound, Diagnostic procedure, Embolism / Thrombosis
Authors:
J. Oppenheimer, R. Mandegaran, B. Kainz, M. P. Heinrich, F. Noor, S. Mischkewitz, A. Ruttloff, P. Klein-Weigel
DOI:
10.26044/ecr2022/C-10357
Conclusion
Machine learning methods can safely aid non-experts in acquiring valid ultrasound images of venous compressions. These can then be externally reviewed for final expert triage and potential diagnosis, resulting in a very high NPV. This ensures that all patients receiving a negative result from the remote expert were in fact DVT-negative. Over 90% of scans taken by a non-expert with the help of AutoDVT were of diagnostic image quality, this increased to 100% with 2-point compression exams excluding the thigh. This study constituted a small dataset with only few positive DVT cases, representing an average rate of positive to negative cases in clinical settings.
In a remote setting, reviewers may be more cautious to rule out DVT, possibly resulting in poor results for specificity and PPV. This technology could also be used as a triage tool to prioritise low risk patients (full compressions of the veins in the relevant anatomical landmarks) and high-risk patients (incomplete compressions of the veins in the relevant anatomical landmarks or insufficient image quality for diagnosis) (see figure 6).
The low-risk patients could avoid long waiting times and come back for a formal ultrasound five to seven days later as per the standard 2-point compression ultrasound protocol and high-risk patients could garner immediate work-up for final diagnosis and treatment [3]. A larger study is planned. Previous research has shown the possibility of machine-learning aided diagnosis by a non-expert [5]. Other research has proposed point-of-care ultrasound (POCUS) by hospital medicine providers instead of radiologists with promising results [6].
The proposed remote triage and diagnostics method allows for fast screening at the point-of-care, without the need for an expert radiologist present, thereby reducing costs, need for patient transportations and possible disease transmission in COVID-settings, without a reduction in patient safety. There is much promise in screening out a large portion of the DVT-negative patients which present with possible symptoms and are sent for rule-out scanning. The sensitivity and NPV of 100% for both reviewers means all patients with a remote exclusion of DVT could be correctly sent home at the point of care and all patients with a DVT were correctly diagnosed without the need for an expert to be at the patient’s bedside.