OC 01.1 - Use of artificial intelligence to guide untrained individuals performing self-ultrasound scans of the knee joint

Saturday, June 22, 2024

13:00 – 13:15 ICT

Room: Ballroom B1

Co-Author(s)

NT

Nima Teimouri

Senior Data Scientist
Novo Nordisk A/S, Søborg, Denmark
Søborg, Sjelland, Denmark
DM

Daniel Michelsanti

Data Scientist
Novo Nordisk A/S, Søborg, Denmark
Søborg, Sjelland, Denmark
PH

Peter Waede Hansen

Senior International Medical Manager
Novo Nordisk A/S, Søborg, Denmark
Søborg, Sjelland, Denmark
SK

Susan Kras

Clinical Liaison Medical Radiation Sciences
Mohawk College, Institute for Applied Health Sciences, McMaster University, Hamilton, Ontario, Canada
Hamilton, Ontario, Canada

Presenting Author(s)

Johannes Roth, MD PhD FRCPC RhMSUS

Chief Pediatric Rheumatology and Rare Diseases
Children’s Hospital of Central Switzerland and University of Luzern
Luzern, Luzern, Switzerland

Co-Author(s)

LB

Lars Bryup

Senior Digital Innovation Consultant
Novo Nordisk A/S, Søborg, Denmark
Søborg, Sjelland, Denmark
MZ

Marek Zak

Senior International Medical Director
Novo Nordisk A/S, Søborg, Denmark
Søborg, Sjelland, Denmark

Background: Recurrent joint bleeds occur frequently in persons with haemophilia, often resulting in arthropathy. The value of ultrasonography (US) in detecting bleeds has been demonstrated; however, access to proficient US operators is limited. This innovative solution proposes the integration of real-time artificial intelligence (AI) assistance, which could empower less experienced physicians and even untrained patients to perform accurate US examinations to acquire US frames of diagnostic quality.

Aims: To investigate the performance of the AI solution in assisting untrained individuals with the scanning procedure, to acquire US frames suitable for synovial recess distension detection, indicating a joint bleed.

Methods: An AI algorithm was developed to provide real-time assessments of US frames collected during the scanning process. Two separate studies (Test I and Test II) were conducted to collect US data. We trained the algorithm with scans of 30 participants (Test I) and validated it on US scans of 6 participants (Test II); (Figure 1). Participants were non-haemophilic and untrained in performing US. Prior to the scan, participants were provided with a video on how to perform a self-scan of the knee. During the scan, they were provided with real-time feedback from the AI algorithm (via a mobile application), to assist finding the correct location and capturing diagnostic-quality US images. To enable algorithm training and validation, images were labelled by a medical expert, based on a verified labelling protocol.

Results: A total of 27,465 (Test I) and 3,105 (Test II) images were acquired via 30 and 6 scans, respectively. Of these, 5,493 and 621 images were labelled by a medical expert. The algorithm correctly classified 449/621 images (72%) from Test II as usable/not usable, demonstrating a sensitivity of 76%, and a specificity of 71%.

Conclusion(s): These data support the feasibility of using real-time AI-assisted software to assist untrained patients performing US self-examinations to acquire diagnostic-quality images.