Prediction model for knee osteoarthritis including clinical, genetic and biochemical risk factors

Abstract

Purpose: Identification of subjects at high risk for knee osteoarthritis (OA) is warranted for biochemical /pharmaceutical research and for prevention/monitoring in the general practice. It is unknown if we are able to distinguish persons who develop knee OA from those who remain free of OA. The objective of this study was to create and compare different risk prediction models including clinical, genetic and/or biochemical risk factors, for knee OA in an elderly population and assess the discriminative value of these models in 4 independent studies. Methods: The prediction model was created in the Rotterdam Study-I (474 incident knee OA cases, 2154 controls of subjects aged 55 years and over) including “questionnaire” and easily obtainable data (for example age, gender, BMI, pain and general health), x-ray data (knee baseline KL score, hand OA, hip OA) and genetic markers. The SNPs chosen to create a genetic risk score are 9 SNPs which were found consistently associated to knee OA in large-scale meta-analyses, including the GDF5, MCF2L and chr7q22 loci. Univariate and multivariate regression models were applied to assess the relationship between the risk factors and incident knee OA. Validation of the model was done in RS-II. Results: The multivariate analysis showed the strongest association(s) with gender (OR 1.69), BMI (OR per sd increase 1.28), hand OA (OR 1.45), knee pain (OR 1.62) and baseline KL score of 1 (OR 6.97). The area-under-the-curve (AUC) is a measure of discrimination between cases and controls with a range from 0.50 (tossing a coin) to 1.00 (perfect prediction). In RS-II (external validation) the AUC for gender, age and BMI in prediction for knee OA was 0.59. Addition of the questionnaire variables or questionnaire variables + genetic score did not change the AUC (AUCs of 0.61 for both models). However, when adding the knee baseline KL score of 0 or 1 to the model the AUC increased to 0.86 (full model). Similar results were observed in the Rotterdam Study-I (internal validation). For uCTXII levels similar results as for the genetic risk score were obtained in a subset of the Rotterdam Study with AUCs of 0.63 (age, gender, BMI + questionnaire), 0.64 (age, gender, BMI, questionnaire, uCTXII) and 0.85 (full model) in RS-II. Validation in Chingford, TwinsUK and the Osteoarthritis Initiative will be assessed in the near future. Conclusions: We showed that the baseline KL score for the knee is the best predictor of future knee OA, next to age, gender and BMI, in an elderly population. We would therefore recommend that radiologists report such findings to general practioners, which is currently not done regularly. “Questionnaire” variables, genetic markers, uCTXII levels and OA at other joint sites do not add much predictive value to age, gender and BMI, at least not in an elderly population.