Dementia progresses slowly from the asymptomatic stage to a fully expressed clinical syndrome over many years. As no effective therapy is currently available, correctly determining whether a person will progress to dementia within the near future has become a public health priority. Nonetheless, it is unknown how to best predict dementia onset. The advent of proteomics offers us an unprecedented opportunity to examine this.

This study included 52,645 non-demented UK Biobank adults (1,417 incident cases; follow-up 14.1 years). Cox-proportional hazard regressions were conducted to estimate the associations of proteins with incident dementia. Receiver operating characteristic analyses and Kaplan-Meier survival curves were performed to evaluate the accuracies and clinical progression of the proteins, respectively. Locally weighted scatterplot smoothing curves were employed to plot the dynamic trajectories.

Of 1,463 plasma proteins, GFAP, NEFL, GDF15, and LTBP2 consistently had the most significant associations with incident all-cause dementia (ACD), Alzheimer's disease (AD), and vascular dementia (VaD) and ranked high in protein importance ordering. Combining GFAP (or GDF15) with demographic data produced desirable predictions for ACD (AUC=0.891) and AD (AUC=0.872) (or VaD (AUC=0.912)). The same parsimonious models yielded AUCs of 0.872, 0.847, and 0.895 for over 10-year ACD, AD, and VaD incidence, respectively, with comparable performance to the full model combining the protein panel and clinical information. Individuals with higher GFAP levels were 2.32 times more likely to develop dementia. GFAP and the novel biomarker LTBP2 were highly specific for dementia prediction. Furthermore, GFAP and NEFL levels began to increase abnormally at least 10 years before dementia diagnosis.