Deep convolutional neural network (DCNN) provides a multivariate framework to detect relevant spatio-oscillatory patterns in the data beyond common mass-univariate statistics. Yet, its practical application is limited due to the low interpretability of the results beyond accuracy. We opted to use DCNN with a minimalistic architecture design and large penalized terms to yield a generalizable and clinically relevant network model, which was trained based on electroencephalography topography data from a primary sample of Parkinson’s disease (PD) patients and healthy controls, with and without medication.
The Jupyter notebook shared here contains the script for building, training, cross-validating, and testing our 3D-CNN to differentiate Parkinson's Disease patients from healthy control based on the spatio-oscialtory pattern.
Method summary. A) The continuous resting-state eyes-open/-closed EEG (rs-EEG) was examined for artifacts and the segments of data with artifacts were flagged. B) The continuous rs-EEG was segmented into 1-second-long trials and the ones containing artifacts were discarded. C) Topographic spectral maps were derived from the first temporal derivatives of the rs-EEG and then normalized to a gamma distribution. D) One topographic map per frequency bin was achieved, followed by stacking matrices into tensors serving as the 3D input volume of the 3D-DCNN. E) Each convolutional layer in the 3D-DCNN architecture is represented by a tesseract. The output of the last convolutional layer was batch normalized and flattened prior to applying the dropout with a rate of 50%. After passing through the dropout layer, the resulting sparse output was fed into a Rectified Linear Unit (ReLU) layer followed by a softmax output layer with 3 neurons.