Brain ERP components predict which individuals progress to Alzheimer's disease and which do not

Abstract

Predicting which individuals will progress to Alzheimer's disease (AD) is important in both clinical and research settings. We used brain Event-Related Potentials (ERPs) obtained in a perceptual/cognitive paradigm with various processing demands to predict which individual Mild Cognitive Impairment (MCI) subjects will develop AD versus which will not. ERP components, including P3, memory "storage" component, and other earlier and later components, were identified and measured by Principal Components Analysis. When measured for particular task conditions, a weighted set of eight ERP component_conditions performed well in discriminant analysis at predicting later AD progression with good accuracy, sensitivity, and specificity. The predictions for most individuals (79%) had high posterior probabilities and were accurate (88%). This method, supported by a cross-validation where the prediction accuracy was 70-78%, features the posterior probability for each individual as a method of determining the likelihood of progression to AD. Empirically obtained prediction accuracies rose to 94% when the computed posterior probabilities for individuals were 0.90 or higher (which was found for 40% of our MCI sample).

Figure 1

The mean brain ERPs to relevant and irrelevant stimuli for the AD(n=12), MCI (Progress(n=15) and Stable(n=28)), and like-age Control(n=12) groups. The ERP data for the MCI individuals were collected approximately 1–2 years before clinical follow-up where progression to AD or stability was determined. For demographic information concerning the ADs and the like-age Controls, see Chapman et al., 2007. The ERPs here were recorded at CZ with reference to linked earlobes. The ERPS shown are from relevant and irrelevant stimulus conditions averaged over all other conditions. For underlying ERP components, see Figure 2 for temporal waveforms (loadings) and Table 3 for group mean component scores.

Figure 2

Temporal waveforms of the ERP components selected by the stepwise discriminant procedure for their predictive power of progression to AD in MCI. Each component is offset here with its baseline shown as a dashed line. In these component waveforms, the metric has been restored by multiplying the loading at each time point by the standard deviation of the data set at the corresponding time point (Chapman and McCrary, 1995). The amplitudes depicted are for a component score of 1.0. The five components appear (top to bottom) in the order they were selected by the discriminant procedure. The discriminant procedure actually selected component_conditions (these were ERP components scores under either relevant or irrelevant stimulus conditions; for more information, see Table 3 and Table 4).

Figure 3

The prediction accuracy for, and number of subjects in, each posterior probability bin based on the development set. Prediction accuracy is the percent of MCI individuals correctly predicted to progress to AD or remain stable. Each of the 43 MCI individuals was placed in a bin by the probability of group membership calculated through the discriminant function (placement was determined by the probability greater than 0.50). The number of subjects is given below the x-axis (e.g. 16/17 = 16 of 17 subjects were correctly classified in this posterior probability bin). Disregarding the bins with low probability (those that were “too close to call”), the prediction accuracy and number of subjects data show 79% of the subjects were predicted with high confidence (in the 0.70 – 1.0 bins) and 88% of these predictions were correct.