Accuracy of analyte signatures in classifying controls and MCI progressors.

Classification accuracy was tested using log10-transformed data. We have used bold font to draw particular attention to the signatures with the best performance (as assessed by MCC) in each comparison (i.e. 10-fold cross-validation, training set or test set). As expected, using all 146 analytes leads to some overfitting on the training set but our signatures that include the APOE analyte perform better in test scenarios.aThe full set of samples contained data on 54 Controls and 163 MCI Progressors.bThe size-matched groups contained data on 54 Controls and 54 MCI Progressors. For each of these datasets, all samples were used for cross-validation, whereas training and test sets were created by dividing datasets into two equal subsets.