A classification modeling approach for determining metabolite signatures in osteoarthritis

Abstract

Multiple factors can help predict knee osteoarthritis (OA) patients from healthy individuals, including age, sex, and BMI, and possibly metabolite levels. Using plasma from individuals with primary OA undergoing total knee replacement and healthy volunteers, we measured lysophosphatidylcholine (lysoPC) and phosphatidylcholine (PC) analogues by metabolomics. Populations were stratified on demographic factors and lysoPC and PC analogue signatures were determined by univariate receiver-operator curve (AUC) analysis. Using signatures, multivariate classification modeling was performed using various algorithms to select the most consistent method as measured by AUC differences between resampled training and test sets. Lists of metabolites indicative of OA [AUC > 0.5] were identified for each stratum. The signature from males age > 50 years old encompassed the majority of identified metabolites, suggesting lysoPCs and PCs are dominant indicators of OA in older males. Principal component regression with logistic regression was the most consistent multivariate classification algorithm tested. Using this algorithm, classification of older males had fair power to classify OA patients from healthy individuals. Thus, individual levels of lysoPC and PC analogues may be indicative of individuals with OA in older populations, particularly males. Our metabolite signature modeling method is likely to increase classification power in validation cohorts.