Background. Mild cognitive impairment (MCI) is associated with an increased risk of developing dementia. Recently published results of the Current Concepts in MCI Conference suggested subclassifications for MCI (MCI-amnestic, MCI-multiple domains slightly impaired, MCI-single nonmemory domain) based on the recognized heterogeneity in the use of the term. These subclassifications have not been empirically validated to date.

Method. A community sample of 1045 dementia-free individuals aged 75 years and over was examined by neuropsychological testing in a three-wave longitudinal study. The prevalences and the predictive validities for the subclassifications of MCI and their modifications (original criteria except for the report of subjective decline in cognitive function) were determined.

Results. The prevalence was 1 to 15% depending on the subset employed. Subjects with a diagnosis of MCI progressed to dementia at a rate of 10 to 55% over 2·6 years, depending on the subset employed. MCI-amnestic achieved the highest positive predictive power (PPP). ROC curves of the subclassifications for MCI indicate that all but one subset for MCI failed to predict dementia (MCI-multiple domains slightly impaired-modified: AUC=0·585, P<0·01, 95% CI, 0·517–0·653). The use of modified criteria for MCI (original criteria except for the report of subjective decline in cognitive function) is associated with a higher diagnostic sensitivity but also with a reduction in diagnostic specificity and PPP.

Conclusions. Modified criteria should be applied if a concept for MCI with a high sensitivity is required and the original criteria (including subjective cognitive complaint) if a concept with high specificity and high PPP is required.