Generalizable electroencephalographic classification of Parkinson’s Disease using deep learning

There is growing interest in using electroencephalography (EEG) and deep learning (DL) to aid in the diagnosis of neurological conditions like Parkinson’s Disease (PD). Many existing DL approaches to classify PD from EEG data cite performance metrics in the high 90% accuracies, but may be grossly overestimating their real-word capabilities due to information-leakage between training and testing data. Our aim was to characterize the potential of deep learning for classifying PD using a conservative training approach with unseen external testing data. We used publicly available resting-state EEG data from patients with PD from two seperate centers (University of New Mexico (n = 54) and University of Iowa (n = 28)) for our training and testing sets, respectively. We implemented a channelwise convolutional neural network and tuned it using a subjectwise cross validation approach. We found that an approach commonly cited in the literature overestimated performance in excess of 20%, while our pipeline more conservatively estimated performance by epoch (accuracy: 69.2%; sensitivity: 66.5%; specificity: 72.2%) and by subject (accuracy: 77.4%, sensitivity: 76.9%, specificity: 77.8%). Moreover, we show that our model generalized well to an unseen and external testing dataset without degradation in performance by epoch (accuracy: 77.2; sensitivity: 83.5%; specificity: 71.0%) and by subject (accuracy: 83.8%, sensitivity: 88.6%, specificity: 79.0%). These results highlight the effect of information leakage and serve as a new benchmark for future generalization of DL approaches to classify PD using EEG data.