Background/Purpose: Juvenile Idiopathic Arthritis (JIA) is the most prevalent chronic rheumatic disease and can result in disability in children. JIA most commonly affects the knee. Despite the widespread occurrence of JIA, its heterogeneous presentation and lack of reliable biomarkers make diagnosis and quantifying treatment response challenging. If untreated, the synovitis in JIA can lead to cartilage degeneration and even bone erosion. To evaluate knee health status non-invasively, we can use contact accelerometers that capture the vibrations generated by internal friction of articulating surfaces during movement, also known as joint acoustic emissions (JAE). In this study, we used JAEs to evaluate children’s knees and distinguish between healthy, inactive JIA, and active JIA.

Methods: We collected knee JAEs from 51 participants while performing flexion/extensions, including 34 subjects with JIA (23 active/11 inactive) and 17 healthy controls. To these JAE signals, we applied filtering and denoising techniques to extract 15 time-frequency audio features to be used as training data for our Logistic Regression machine learning (ML) classification model. This classifier was implemented in two stages, where at each we obtained a knee health score that indicates the predicted class. The first stage distinguished between healthy and JIA subjects with active and inactive JIA grouped together. The second stage then took only the subjects identified as JIA and classified them as either active or inactive. Hence, our 2-stage classifier was used to evaluate not only presence of, but also severity of JIA.

Results: We validated our 2-stage classifier through leave-one-subject-out cross-validation and obtained an overall 3-class accuracy of 76% when differentiating between healthy, inactive JIA, and active JIA subjects. We obtained a Stage 1 accuracy of 78% and a Stage 2 accuracy of 81%. The least accurate class was inactive JIA, which had both the lower number of sample subjects for training (11) and has the most uncertain label since designation of knee arthritis as “inactive” is primarily clinical and thus is subject to variability. Notice error compounding from both stages results in a lower overall accuracy since second stage depends on first stage predictions, hence, with further ML optimization, we see potential for improved diagnostic accuracy.

Conclusion: These results support the use of JAEs as a novel non-invasive digital biomarker for knee joint health assessment. More accurate characterization of synovitis by magnetic resonance imaging can facilitate improved accuracy of our training dataset which in turn could allow us to differentiate between different levels of synovial inflammation. We have shown that close collaboration between clinicians and bioengineers can result in personalized assessment and thereby improving management of arthritis in children.

« Back to ACR Convergence 2021

ACR Meeting Abstracts - https://acrabstracts.org/abstract/joint-acoustic-emissions-as-a-biomarker-to-differentiate-between-active-and-inactive-juvenile-idiopathic-arthritis-via-2-stage-machine-learning-classifier/