EXTERNAL VALIDATION OF AN ARTIFICIAL INTELLIGENCE TOOL FOR RADIOGRAPHIC KNEE OSTEOARTHRITIS CLASSIFICATION: A MULTI-CENTER, RETROSPECTIVE DIAGNOSTIC COHORT STUDY

• Published in: Osteoarthritis imaging 2023, Vol.3, p.100115, Article 100115
• Main Authors: Brejnebøl, M.W., Lenskjold, A., Ziegeler, K., Ruitenbeek, H.C., Visser, J.J., Nybing, J.U., Hermann, K.G.A., Oei, E.H.G., Boesen, M.
• Format: Article
• Language: English
• ISSN: 2772-6541; 2772-6541
• DOI: 10.1016/j.ostima.2023.100115

In routine clinical practice, radiography of the weight-bearing knee is frequently obtained to support the OA diagnostic process. The reading time of radiographs is short, but the volume of studies is large, resulting in a substantial radiologist workload. Artificial Intelligence (AI) tools for radiographic knee OA severity classification have been commercially available for some time. However, the generalizability of these tools across countries remains limited. The overall objective of this study was to evaluate a commercially available AI tool for radiographic knee OA classification across a heterogenous clinical dataset. Our specific objectives were: 1-Investigate the overall diagnostic performance of the AI tool for KLG classification. 2-Investigate the sub-group performance for posterior-anterior (PA) vs. anterior-posterior (AP) projection radiographs. This was a multi-center, retrospective diagnostic test accuracy study. We consecutively included seventy-five cases from the production Picture Archiving and Communications System (PACS) of each of the three EU-based study sites, yielding a sample size of n = 225. Inclusion criteria were age >= 20 years, clinical suspicion of knee OA, at least one weight-bearing PA or AP projection knee radiograph and one lateral knee radiograph of the symptomatic knee. Exclusion criteria were knee arthroplasty or other foreign objects near the knee joint, and the study was from an already included patient. The three study sites (Bispebjerg and Frederiksberg Hospital, Copenhagen, Denmark, Erasmus Medical Center, Rotterdam, the Netherlands, and Charité Universitätsmedizin, Berlin, Germany) routinely obtained their frontal knee radiographs differently: one site used PA, one AP, and one full-leg stitched AP where the knee joint was not necessarily centered to the x-ray beam. Only one symptomatic knee was used for each patient. The reference standard was established based on the independent scoring of three professors in musculoskeletal radiology, K-GAH, EHGO, and MB (one from each site). In case of discrepancies, the reference value was the majority vote of the three or by consensus meeting when KLGs differed by two or more. A knee was discarded if two or more scored an image as inadequate for diagnostic use (equivalent to recalling the patient for a new radiograph). MWB performed the AI tool (RBknee™ v2.1) analysis on a local machine. Weighted accuracy for ordinal values (OWA) was used for the ordinal scores, and regular