Background: Cognitive impairment precedes the diagnosis of Alzheimer's disease. It is unclear which psychometric measures predict dementia, and what cut-off points should be used. Replicable cognitive measures to provide information about differential diagnosis and prognosis would be clinically useful. Aims: In a prospective cohort study we investigated which measures distinguish between individuals with amnestic mild cognitive impairment (aMCI) that converts to dementia and those whose impairment does not, and which combination of measures best predicts the fate of people with aMCI. Method: Forty-four participants with aMCI underwent extensive neuropsychological assessment at baseline and annually thereafter for an average of 4 years. Differences in baseline cognitive performance of participants who were converters and non-converters to clinically diagnosed dementia were analysed. Classification accuracy was estimated by sensitivity, specificity, positive and negative predictive values and using logistic regression. Results: Forty-one percent of participants had progressed to dementia by the end of study, with a mean annual conversion rate of 11%. Most (63%) showed persisting or progressive cognitive impairment, irrespective of diagnosis. The Addenbrooke's Cognitive Examination together with the discrimination index of the Hopkins Verbal Learning Test - Revised (but none of the demographic indices) differentiated the participants who were converters from the nonconverters at baseline with 74% accuracy. Conclusions: Targeted neuropsychological assessment, beyond simple cognitive screening, could be used in clinical practice to provide individuals with aMCI with prognostic information and aid selective early initiation of monitoring and treatment among those who progress towards a clinically diagnosable dementia.