Background: Age is the strongest risk factor for dementia and there is considerable interest in identifying scalable, blood-based biomarkers in predicting dementia. We examined the role of midlife serum metabolites using a machine learning approach and determined whether the selected metabolites improved prediction accuracy beyond the effect of age. Methods: Five thousand three hundred seventy-four participants from the Whitehall II study, mean age 55.8 (standard deviation (SD) 6.0) years in 1997–1999 when 233 metabolites were quantified using nuclear magnetic resonance metabolomics. Participants were followed for a median 21.0 (IQR 20.4, 21.7) years for clinically-diagnosed dementia (N=329). Elastic net penalized Cox regression with 100 repetitions of nested cross-validation was used to select models that improved prediction accuracy for incident dementia compared to an age-only model. Risk scores reflecting the frequency with which predictors appeared in the selected models were constructed, and their predictive accuracy was examined using Royston's R2, Akaike's information criterion, sensitivity, specificity, C-statistic and calibration. Results: Sixteen of the 100 models had a better c-statistic compared to an age-only model and 15 metabolites were selected at least once in all 16 models with glucose present in all models. Five risk scores, reflecting the frequency of selection of metabolites, and a 1-SD increment in all five risk scores was associated with higher dementia risk (HR between 3.13 and 3.26). Three of these, constituted of 4, 5 and 15 metabolites, had better prediction accuracy (c-statistic from 0.788 to 0.796) compared to an age-only model (c-statistic 0.780), all p<0.05. Conclusions: Although there was robust evidence for the role of glucose in dementia, metabolites measured in midlife made only a modest contribution to dementia prediction once age was taken into account.