EEG machine learning for accurate detection of cholinergic intervention and Alzheimer’s disease

Figures

• 

  Figure 1. Spectral and temporal correlation biomarkers exhibit sensitivity to scopolamine administration. (a) EEG of a subject in the baseline (blue) and scopolamine (red) condition. (b) Grand average normalized power spectra indicate large effects of scopolamine, most notably a reduction of power in the alpha and beta bands, and an increase of delta and theta power. (c) Oscillation dynamics were studied by extracting the amplitude envelope from band-pass filtered data (e.g., the alpha band, black) using the Hilbert transform (blue, red) and a median-amplitude threshold to determine the onset and offset of a burst. (d) A cumulative probability distribution of all oscillation bursts revealed a tendency towards longer alpha bursts in the baseline condition. (e) Amplitude envelopes of beta oscillations (13–30 Hz) suggest a more complex temporal structure in the peak scopolamine (red) than in the baseline (blue) condition on time scales of seconds to tens of seconds. (f) The long-time scale differences in beta oscillations are reflected in the grand average DFA showing larger scaling exponents for peak scopolamine (red) than for the baseline recording (blue). All figures were based on the…

• 

  Figure 2. Scopolamine affects many characteristics of the EEG. Time dependence of different EEG biomarkers (columns) and frequency bands (rows) for placebo (black) and scopolamine (red). All biomarkers are shown as averages over the 2 channels, except for relative power, for which the Fz-Cz and Pz-Oz channels are shown separately, because the effects in the delta and beta bands were opposite for the two derivatives. Sixteen biomarkers were significantly affected by scopolamine, with the peak effect occurring 1.5 h after administration. The values plotted are group means and standard errors of the mean computed for the within-subject design67. Significance levels: * denotes p < 0.05, **p < 10−5, ***p < 10−10; Bonferroni corrected for multiple comparisons.

• 

  Figure 3. Enhanced detection of scopolamine-induced EEG changes using machine learning. (a) Classification performance increased with the number of features included in the integrated index. (b) All of the biomarkers selected by elastic net logistic regression for inclusion in the mAChR index differed significantly between baseline and peak scopolamine. Biomarkers are ordered by their absolute weights, decreasing clockwise from the top. Weights (β) are listed next to each biomarker in the legend (PO denotes Pz-Oz and FC denotes Fz-Cz). The values plotted on the spider plot are the z-score group means and standard error of the mean, normalized to [0, 1] by subtracting the minimum across all biomarkers and dividing with the largest range present (i.e., the difference between the minimum and maximum value found for the biomarkers with the largest difference). (c) The mAChR index was more sensitive to the scopolamine (Scp) intervention than relative delta power. The plot shows z-scored biomarker values per subject recording. Singled-out symbols represent median values per group with standard error bars. The dashed line indicates the threshold of the classifier to predict the recordings as a…

• 

  Figure 4. The mAChR index generalizes to a new cohort of subjects. (a) The mAChR index also generalizes to a cohort of healthy elderly subjects receiving the same scopolamine intervention (Trial 3), with validation accuracy 87%, sensitivity 83%, specificity 91% and precision 91%. (b) Time dependence curves demonstrate the generalizability of the index at all the time points. The mAChR index for placebo (black) and scopolamine (red) is shown for Trials 1&2 used for developing the index and for independent data from Trial 3 (in the latter there was no placebo condition). The values plotted are group means and standard errors of the mean computed for the within-subject design67.

• 

  Figure 5. The Alzheimer’s index captures scopolamine-induced effects and validates scopolamine as a model of AD pathophysiology. (a) Illustration of the twelve biomarkers composing the AD index. Several of these biomarkers also compose the scopolamine index, with the same directionality of change in both the scopolamine-induced cognitive impairment and Alzheimer’s disease. The values plotted are as explained in Fig. 3B. (b) AD index separates healthy elderly from Alzheimer’s disease patients with high precision. (c) The mAChR index discriminates the healthy elderly and AD patients albeit less accurately than the AD index. (d) Validating the AD index on the scopolamine data gives much better discrimination.