AbstractBackgroundMild cognitive impairment (MCI) is the transitional zone between normal aging and Alzheimer’s dementia (AD), representing a group of subjects with higher risk of conversion to AD (Petersen, 2010). However, as accumulating evidence shows that there are a non‐negligible number of MCI subjects who revert to be cognitive normal (CN), additional measures are needed for better characterizations on the MCI groups (Petersen, 2014). Thus, the purpose of this study is to explore the structural covariance (SC) feature of subjects with different longitudinal clinical status changes.MethodParticipants were from the Beijing Aging Brain Rejuvenation Initiative (BABRI), with two clinical cognitive assessments and the baseline high‐resolution structural MRI data. And four groups were defined, the stable CN (sCN, n = 99), CN progressing to MCI (pCN, n = 23), stable MCI (sMCI, n = 29), and MCI reverting to CN (rMCI, n = 33), and the mean age and follow‐ups interval for CN and MCI subjects were 64.74 and 66.63, 2.46 and 1.45 years, respectively. And the seed‐based multivariate method (Alexander‐Bloch, 2013) was applied to identify the SC of default‐mode network (scDN), frontoparietal network (scFN) and hippocampal network (scHN), of which the covariance scores were calculated. To obtain the SC templates of three networks, 69 sCN were randomly selected, and the rest 30 sCN, with the other three groups, were included in establishing prediction model of clinical status changes based on network SC scores, by using the support vector machine method and leave‐one‐out cross‐validation. Moreover, baseline cognition was also added in the model.ResultFor classification between sCN and sMCI, the scHN score presented the highest accuracy (area under curve (AUC) = 0.89), and it was the scDN score showing the best performance in classifying sMCI and rMCI (AUC = 0.82), with the scFN score outperforming others in distinguishing pCN from sCN (AUC = 0.68). And when baseline cognition was also included, prediction accuracies of clinical changes were improved (AUC = 0.84 for sMCI vs. rMCI, AUC = 0.83 for sCN vs. pCN).ConclusionTaking together, our findings indicated the potential of network structural covariance in classifying MCI subjects with different risk of conversion.