Utilization of resting-state electroencephalography spectral power in convolutional neural networks for classification of primary progressive aphasia.

We investigated relative power spectral density (PSD) in primary progressive aphasia (PPA) in delta, theta, alpha, and beta frequency bands in eyes open and closed resting-state electroencephalography (EEG). Our aims were to assess whether discernible differences could be observed between each PPA variant and to determine the utility of PSD for PPA classification when used as input to a convolutional neural network (CNN). Findings in the current study were similar to previous studies in logopenic PPA, with a significant increase in relative PSD in delta and theta bands and a significant reduction in the beta band (consistent with oscillatory slowing). We did not observe a significant increase in power for lower frequency bands or a reduction of power in higher frequency bands for semantic or nonfluent PPA, in contrast to what has been previously reported. In semantic PPA, evidence pointed to oscillatory speeding, not the slowing that was previously reported in a single-case study. In nonfluent PPA, spectral power fell between logopenic and semantic PPA, suggesting there is oscillatory slowing but to a lesser extent than logopenic PPA. The CNN was relatively successful in distinguishing PPA from healthy controls (F1 = 0.851). The CNN did not perform as well on four-way classification (lvPPA, svPPA, nfvPPA, controls; F1 = 0.586) but was significantly above chance. These results are promising and suggest that resting-state EEG may prove useful as a biomarker for PPA diagnosis. Potential factors underlying the differences between the findings of the current study and previous work are discussed.