Fig. 1: Study workflow.

The F.ACE longitudinal cohort served as the discovery cohort, where people diagnosed with MCI were followed up to monitor dementia conversion. Plasma proteomics were generated from the collected plasma samples at baseline. Proteomic data from the F.ACE discovery cohort underwent quality control (QC), normalization, and marginal screening within a predefined training set. A separate testing set was held out for evaluating selected features. Feature selection was performed using Lasso, RSF minimal depth, and recursive feature elimination (RFE), followed by model training using survival machine learning methods: survival gradient boosting (SGB), random survival forests (RSF), and survival support vector machines (sSVM). External validation was performed using two datasets from the EMIF-AD MBD cohort.