Background/Purpose:

MRI-detected cartilage loss is the main structural outcome measure in large studies of knee OA. Definition of subjects at high risk for cartilage loss is important as this subgroup is likely to benefit most from interventional efforts and will potentially reduce subject numbers and duration of clinical trials. A simple stratification strategy is needed for pre-selecting subjects at high risk. We assessed a set of potential factors in regard to their predictive capability of cartilage loss over 30 months.

Methods:

The Multicenter Osteoarthritis (MOST) Study is a longitudinal study of subjects with knee OA or at risk of OA. 1.0 T MRI was performed at baseline and 30 months follow-up in subjects with radiographic knee OA at baseline. MRIs were assessed according to the WORMS scoring system including within-grade assessment. Altogether 10 tibiofemoral subregions were scored. Cartilage loss was defined as at least within-grade increase in cartilage score in >1, ≥2, or ≥3 subregions. Analyses were repeated with cartilage loss defined as full grade increase. Six predictors that are commonly acquired in screening efforts, i.e., gender, age, BMI, Kellgren Lawrence (K-L) grade (0-1, 2, 3-4), joint effusion (0-1, 2-3) and knee pain (pain on most days during the past 30 days or maximal WOMAC pain score as 0, 1-2, 3-4) were assessed using a 10-fold cross-validation method. The cross-validation was repeated 10 times, and averaged receiver operating characteristics (ROC) curve, i.e., C-statistic, was calculated as a measure of the overall performance. Age and sex were forced in all models. We plotted the ROC curves for each variable selected in the final model according to cross-validation method.

Results:

Of 544 knees randomly selected from the progression cohort, risk of at least within-grade cartilage loss occurring in ≥1, or ≥2, or ≥3 subregions was 53.6%, 31.6%, and 18.1%, respectively. The model containing age, sex, K-L grade, effusion, and knee pain provided the highest prediction accuracy (C-statistic 0.533) for any cartilage loss occurring in at least one subregion. The best models for prediction of at least within-grade cartilage loss of ≥2 or ≥3 included the same variables, i.e., age, sex, K-L grade and knee pain (C-statistics=0.561 for ≥2, and 0.571 for ≥3 subregions). Risk of full-grade cartilage loss in ≥1, ≥2, or ≥3 subregions was 38.8%, 18.8% and 9.7%, respectively. The same variables were included for predicting cartilage loss of a full grade in ≥1, or ≥2, or ≥3 subregions as those for at least within-grade loss. The corresponding C-statistics were 0.559, 0.599, and 0.615 (Figure 1), respectively.

Conclusion:

Among knees with radiographic OA, age, sex, K-L grade and knee pain only had the moderate capability of predicting cartilage loss. Of them the strongest predictors were baseline K-L grade, followed by knee pain. The prediction accuracy of this model needs to be further validated using large databases from other populations.