Context Electroencephalography (EEG) is a complex signal and can require several years of training, as well as advanced signal processing and feature extraction methodologies to be correctly interpreted. Recently, deep learning (DL) has shown great promise in helping make sense of EEG signals due to its capacity to learn good feature representations from raw data. Whether DL truly presents advantages as compared to more traditional EEG processing approaches, however, remains an open question. Objective In this work, we review 154 papers that apply DL to EEG, published between January 2010 and July 2018, and spanning different application domains such as epilepsy, sleep, brain-computer interfacing, and cognitive and affective monitoring. We extract trends and highlight interesting approaches from this large body of literature in order to inform future research and formulate recommendations. Methods Major databases spanning the fields of science and engineering were queried to identify relevant studies published in scientific journals, conferences, and electronic preprint repositories. Various data items were extracted for each study pertaining to (1) the data, (2) the preprocessing methodology, (3) the DL design choices, (4) the results, and (5) the reproducibility of the experiments. These items were then analyzed one by one to uncover trends. Results Our analysis reveals that the amount of EEG data used across studies varies from less than ten minutes to thousands of hours, while the number of samples seen during training by a network varies from a few dozens to several millions, depending on how epochs are extracted. Interestingly, we saw that more than half the studies used publicly available data and that there has also been a clear shift from intra-subject to inter-subject approaches over the last few years. About [Formula: see text] of the studies used convolutional neural networks (CNNs), while [Formula: see text] used recurrent neural networks (RNNs), most often with a total of 3-10 layers. Moreover, almost one-half of the studies trained their models on raw or preprocessed EEG time series. Finally, the median gain in accuracy of DL approaches over traditional baselines was [Formula: see text] across all relevant studies. More importantly, however, we noticed studies often suffer from poor reproducibility: a majority of papers would be hard or impossible to reproduce given the unavailability of their data and code. Significance To help the community progress and share work more effectively, we provide a list of recommendations for future studies and emphasize the need for more reproducible research. We also make our summary table of DL and EEG papers available and invite authors of published work to contribute to it directly. A planned follow-up to this work will be an online public benchmarking portal listing reproducible results.

https://iopscience.iop.org/article/10.1088/1741-2552/ab260c/pdf

Deep learning-based electroencephalography analysis: a systematic review.

Electrophysiological signals exhibit both periodic and aperiodic properties. Periodic oscillations have been linked to numerous physiological, cognitive, behavioral and disease states. Emerging evidence demonstrates that the aperiodic component has putative physiological interpretations and that it dynamically changes with age, task demands and cognitive states. Electrophysiological neural activity is typically analyzed using canonically defined frequency bands, without consideration of the aperiodic (1/f-like) component. We show that standard analytic approaches can conflate periodic parameters (center frequency, power, bandwidth) with aperiodic ones (offset, exponent), compromising physiological interpretations. To overcome these limitations, we introduce an algorithm to parameterize neural power spectra as a combination of an aperiodic component and putative periodic oscillatory peaks. This algorithm requires no a priori specification of frequency bands. We validate this algorithm on simulated data, and demonstrate how it can be used in applications ranging from analyzing age-related changes in working memory to large-scale data exploration and analysis.

Parameterizing neural power spectra into periodic and aperiodic components.

Frontal Cortex and the Hierarchical Control of Behavior.

https://www.cell.com/trends/neurosciences/pdf/S0166-2236(17)30024-3.pdf

Where Does EEG Come From and What Does It Mean

Published in final edited form as:
Nat Neurosci. 2019 May ; 22(5): 820–827. doi:10.1038/s41593-019-0371-x.

/pdf/working-memory-revived-in-older-adults-by-synchronizing-15w4y9qje0.pdf

Working memory revived in older adults by synchronizing rhythmic brain circuits.

Brain Oscillations and the Importance of Waveform Shape

Oscillations in neural activity play a critical role in neural computation and communication. There is intriguing new evidence that the nonsinusoidal features of the oscillatory waveforms may inform underlying physiological and pathophysiological characteristics. Time-domain waveform analysis approaches stand in contrast to traditional Fourier-based methods, which alter or destroy subtle waveform features. Recently, it has been shown that the waveform features of oscillatory beta (13-30 Hz) events, a prominent motor cortical oscillation, may reflect near-synchronous excitatory synaptic inputs onto cortical pyramidal neurons. Here we analyze data from invasive human primary motor cortex (M1) recordings from patients with Parkinson's disease (PD) implanted with a deep brain stimulator (DBS) to test the hypothesis that the beta waveform becomes less sharp with DBS, suggesting that M1 input synchrony may be decreased. We find that, in PD, M1 beta oscillations have sharp, asymmetric, nonsinusoidal features, specifically asymmetries in the ratio between the sharpness of the beta peaks compared with the troughs. This waveform feature is nearly perfectly correlated with beta-high gamma phase-amplitude coupling (r = 0.94), a neural index previously shown to track PD-related motor deficit. Our results suggest that the pathophysiological beta generator is altered by DBS, smoothing out the beta waveform. This has implications not only for the interpretation of the physiological mechanism by which DBS reduces PD-related motor symptoms, but more broadly for our analytic toolkit in general. That is, the often-overlooked time-domain features of oscillatory waveforms may carry critical physiological information about neural processes and dynamics.SIGNIFICANCE STATEMENT To better understand the neural basis of cognition and disease, we need to understand how groups of neurons interact to communicate with one another. For example, there is evidence that parkinsonian bradykinesia and rigidity may arise from an oversynchronization of afferents to the motor cortex, and that these symptoms are treatable using deep brain stimulation. Here we show that the waveform shape of beta (13-30 Hz) oscillations, which may reflect input synchrony onto the cortex, is altered by deep brain stimulation. This suggests that mechanistic inferences regarding physiological and pathophysiological neural communication may be made from the temporal dynamics of oscillatory waveform shape.

https://www.jneurosci.org/content/jneuro/37/18/4830.full.pdf

Nonsinusoidal beta oscillations reflect cortical pathophysiology in Parkinson's disease

We introduce a fully documented, open-source Python package, bycycle, for analyzing neural oscillations on a cycle-by-cycle basis. This approach is complementary to traditional Fourier- and Hilbert...

Cycle-by-cycle analysis of neural oscillations

Neural activity in the β frequency range (13-30 Hz) is excessively synchronized in Parkinson's disease (PD). Previous work using invasive intracranial recordings and non-invasive scalp electroencephalography (EEG) has shown that correlations between β phase and broad-band γ (>50 Hz) amplitude [i.e., phase amplitude coupling (PAC)] are elevated in PD, perhaps a reflection of this synchrony. Recently, it has also been shown, in invasive human recordings, that non-sinusoidal features of β oscillation shape also characterize PD. Here, we show that these features of β waveform shape also distinguish PD patients on and off medication using non-invasive recordings in a dataset of 15 PD patients with resting scalp EEG. Specifically, β oscillations over sensorimotor electrodes in PD patients off medication had greater sharpness asymmetry and steepness asymmetry than on medication (sign rank, p < 0.02, corrected). We also showed that β oscillations over sensorimotor cortex most often had a canonical shape, and that using this prototypical shape as an inclusion criteria increased the effect size of our findings. Together, our findings suggest that novel ways of measuring β synchrony that incorporate waveform shape could improve detection of PD pathophysiology in non-invasive recordings. Moreover, they motivate the consideration of waveform shape in future EEG studies.

/pdf/characteristics-of-waveform-shape-in-parkinson-s-disease-178o8r8yqj.pdf

Characteristics of Waveform Shape in Parkinson's Disease Detected with Scalp Electroencephalography.

Neural oscillations are widely studied using methods based on the Fourier transform, which models data as sums of sinusoids. For decades these Fourier-based approaches have successfully uncovered links between oscillations and cognition or disease. However, because of the fundamental sinusoidal basis, these methods might not fully capture neural oscillatory dynamics, because neural data are both nonsinusoidal and non-stationary. Here, we present a new analysis framework, complementary to Fourier analysis, that quantifies cycle-by-cycle time-domain features. For each cycle, the amplitude, period, and waveform symmetry are measured, the latter of which is missed using conventional approaches. Additionally, oscillatory bursts are algorithmically identified, allowing us to investigate the variability of oscillatory features within and between bursts. This approach is validated on simulated noisy signals with oscillatory bursts and outperforms conventional metrics. Further, these methods are applied to real data, including hippocampal theta, motor cortical beta, and visual cortical alpha, and can differentiate behavioral conditions.

https://www.biorxiv.org/content/biorxiv/early/2018/04/16/302000.full.pdf

Scott R. Cole

Papers

Brain Oscillations and the Importance of Waveform Shape

Nonsinusoidal beta oscillations reflect cortical pathophysiology in Parkinson's disease

Cycle-by-cycle analysis of neural oscillations

Characteristics of Waveform Shape in Parkinson's Disease Detected with Scalp Electroencephalography.

Cycle-by-cycle analysis of neural oscillations