Session Type: ACR Poster Session C
Session Time: 9:00AM-11:00AM
Currently there are a large number of postulated predictors for OA progression, proposed on the basis of relatively small studies (100’s), and measurement approaches involving large error margins. Supervised machine learning techniques, allows the study of key pathologies without prior assumptions, and accurate quantified segmentation. Our aim was to evaluate predictors of cartilage loss in the entire OAI dataset.
Automated cartilage segmentation was assessed for repeatability and accuracy and then applied to all knees from the OAI with baseline, 1 and 2-year images. Shape models were used to measure cartilage thickness in a consistent anatomical region of the medial tibiofemoral joint.
Multilevel modelling was used to consider the effects of previous postulated covariates at knee-level and subject-level at baseline and in longitudinal change. Gender, KL grade, race, knee alignment, height and weight, age, pain, physical activity, NSAID use, previous surgery, smoking and systolic blood pressure were used as covariates.
The segmentation method showed excellent accuracy; miss-n-out validation in a training set of 287 knees with varying radiographic OA, showed a mean point-to-surface accuracy of 0.12 mm, mean 95% error of 0.37mm (less than one voxel edge). 9,254 knees were included; mean age was 61.2 (SD 9.2); 58.5% were female. 24.8% used NSAIDs, 47.2% were smokers, 22.7% had previous surgery.
Gender and KL grade explained almost all the variability in cartilage thickness at baseline; for example, the adjusted pseudo-r2 of femur and tibia models using only gender and KL grade were 38% and 40%, the full models were 40% and 43%, coefficients are provided in Table 1. A similar pattern was found in longitudinal analysis (Table 2); gender and KL grade accounted for almost all the variability in change in thickness.
This is the first study to characterise the relationship of cartilage thickness, and its change, using a very large dataset. Two covariates explained almost all the variance in thickness: gender and radiographic disease, previously described factors such as weight and alignment (nor NSAIDS and smoking) had surprisingly little effect. This changes our concept of osteoarthritis progression, and potential interventions.
To cite this abstract in AMA style:Bowes MA, Alabas O, Guillard G, Vincent GR, Brett A, Conaghan PG. Contrary to Current Understanding of Osteoarthritis Pathogenesis, Only Gender and Radiographic Status Predict Progression: A Machine Learning Study of 9,254 Knees from the Osteoarthritis Initiative [abstract]. Arthritis Rheumatol. 2017; 69 (suppl 10). https://acrabstracts.org/abstract/contrary-to-current-understanding-of-osteoarthritis-pathogenesis-only-gender-and-radiographic-status-predict-progression-a-machine-learning-study-of-9254-knees-from-the-osteoarthritis-initiative/. Accessed November 18, 2019.
« Back to 2017 ACR/ARHP Annual Meeting
ACR Meeting Abstracts - https://acrabstracts.org/abstract/contrary-to-current-understanding-of-osteoarthritis-pathogenesis-only-gender-and-radiographic-status-predict-progression-a-machine-learning-study-of-9254-knees-from-the-osteoarthritis-initiative/