PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep learning-based electroencephalography analysis: a systematic review

Yannick Roy, Hubert Banville, Isabela Albuquerque, Alexandre Gramfort, Tiago H. Falk, Jocelyn Faubert

arXiv:1901.05498v2cs.LGeess.SPstat.ML

TL;DR

EEG’s low signal-to-noise ratio and non-stationarity complicate reliable analysis, leaving open whether deep learning improves on traditional approaches. This review of 156 studies found a median 5.4% accuracy gain for deep learning over baselines.

  • Problem

    EEG’s low signal-to-noise ratio and non-stationarity make effective analysis difficult, especially when limited data must generalize across recording times.

  • Method

    The authors systematically reviewed 156 deep-learning EEG studies published from 2010 to July 2018, analyzing data, processing, models, results, and reproducibility.

  • Results

    5.4% was the median accuracy gain of deep learning over traditional baselines across relevant studies.

  • Takeaways & Limitations

    The review recommends clearer reporting, shared data and code, existing datasets, and strong baselines to improve reproducibility and fair comparison.

  • Takeaways & Limitations

    To keep the review manageable, the authors omitted various factors affecting deep-learning EEG performance and excluded some collected data items from analysis.

Abstract

from arXiv · show

Electroencephalography (EEG) is a complex signal and can require several years of training 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. In this work, we review 156 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 in order to inform future research and formulate recommendations. 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. Our analysis reveals that the amount of EEG data used across studies varies from less than ten minutes to thousands of hours. As for the model, 40% of the studies used convolutional neural networks (CNNs), while 14% used recurrent neural networks (RNNs), most often with a total of 3 to 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 5.4% 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. To help the field progress, we provide a list of recommendations for future studies and we make our summary table of DL and EEG papers available and invite the community to contribute.

1 Introduction

EEG offers millisecond-scale temporal resolution and broad clinical and brain-computer-interface applications, but its noisy, non-stationary, and highly subject-variable signals complicate robust analysis. Deep learning may enable more flexible, end-to-end EEG processing, motivating this systematic review of the field.

  • EEG background: EEG captures brain electrical activity with excellent millisecond-scale temporal resolution but low spatial resolution because tissues smear signals between sources and sensors.EEG channels are often highly correlated spatially as a result.
  • EEG applications: EEG supports clinical applications such as sleep and epilepsy analysis, monitoring of brain-related conditions, and brain-computer interfaces.BCIs bypass natural output pathways to let brain activity directly affect the user’s environment.
  • EEG limitations: Low signal-to-noise ratio, non-stationarity, and high inter-subject variability hinder EEG processing and generalization across times and individuals.Artifacts can match or exceed brain-signal amplitude, temporal shifts can impair same-subject generalization, and physiological differences can severely affect cross-subject model performance.
  • Existing processing: Domain-specific pipelines commonly clean, extract features, and classify EEG data, but they reduce the flexibility and generalization capability of EEG technologies.Existing approaches include Riemannian geometry-based and adaptive classifiers, while tasks such as sleep scoring remain time-consuming when performed manually.
  • Deep learning motivation: Deep learning could simplify EEG processing by learning preprocessing, feature extraction, and classification end to end from raw or minimally preprocessed data.Its hierarchical representations may reduce domain-specific processing and produce features more expressive than those engineered by humans.
  • Review rationale: EEG’s relatively small, costly, and privacy-constrained datasets provide reasons for skepticism about applying deep learning to EEG.The review therefore analyzes many recent publications to summarize the state of DL-based EEG processing for traditional EEG and DL researchers.

2 Methods

The review targeted English journal, conference, and preprint publications from January 2010 through July 2018, using database searches, reference screening, and staged eligibility assessment. For each included article, the authors extracted around 70 data items spanning five categories.

  • Search strategy: January 2010–July 2018 was the publication window, with PubMed, Google Scholar, and arXiv searched and reference lists scanned for additional papers.The databases were last queried on July 2, 2018.
  • Search strategy: 18 title-and-abstract search terms covered EEG, deep learning, neural-network architectures, and related models such as CNNs, RNNs, GANs, autoencoders, and DBNs.The terms were combined using logical operators, though the supplied passage truncates the search-string description.
  • Study selection: Eligibility was assessed sequentially through title screening, abstract review when necessary, and full-text rejection of papers misaligned with the inclusion and exclusion criteria.This staged process was applied during selection and data collection.
  • Study selection: Non-peer-reviewed preprints were included to capture timely and unconventional research, increase diversity, and reduce possible peer-review bias.The passage specifically identifies arXiv electronic preprints as an example.
  • Data extraction: Around 70 data items per article were extracted using an iteratively refined table, with each paper initially reviewed by one author and reassessed by a second when needed.The extraction items covered five categories, and the table was informed by previous reviews and the authors’ field knowledge.

3 Results … 3.4 EEG processing

The review analyzed 156 DL-EEG papers spanning diverse application domains, datasets, recording parameters, augmentation strategies, and EEG-processing choices. Across these studies, preprocessing was common, artifact handling varied, and raw EEG inputs were nearly as prevalent as hand-engineered features.

  • 3 Results: 156 papers were selected for analysis after 553 database results, 49 additional reference-identified papers, and 446 exclusions.
  • 3.1 Origin of the selected studies: 49 journal papers, 58 conference and workshop papers, 48 preprints, and 1 journal supplement met the inclusion criteria.
  • 3.2 Domains: 86% of studies focused on EEG classification, especially sleep staging, seizure detection or prediction, BCIs, and cognitive or affective monitoring.Around 9% addressed processing-tool improvements, while the remaining 5% covered other applications.
  • 3.3 Data: EEG data quantities varied substantially across domains, from many-hour recordings in sleep and epilepsy to a few hours or minutes in affective and cognitive monitoring.Example counts also varied with windowing: 4,050,000 examples came from 1080 minutes using 250 ms windows with 234 ms overlap.
  • 3.3 Data: Half of datasets contained fewer than 13 subjects, whereas some studies used at least 250 subjects and others used 10,000 or 16,000 subjects.The review identifies combining subjects or datasets as a potential way to learn common features and improve generalization.
  • 3.3 Data: EEG recordings used 1 to 256 electrodes, with half of studies using 8 to 62 electrodes, while sampling rates mostly ranged from 100 to 1000 Hz.Around 50% used sampling rates of 250 Hz or less, and the highest reported rate was 5000 Hz.
  • 3.3 Data: Three papers explicitly studied data augmentation, while 30 additional papers used augmentation but only a handful investigated its performance impact.In one study, overlapping-window augmentation increased accuracy from 60.91% to 74.33%.
  • 3.4 EEG processing: 72% of studies used at least one preprocessing method, 46% omitted artifact handling, 24% used it, and 31% did not report whether it was needed.For model inputs, 49% used only raw EEG, 48% used hand-engineered features, and 3% did not specify the input type.

3.5 Deep learning methodology · 3.6 Inspection of trained models

The review found CNNs dominant in DL-EEG architectures, usually with shallow-to-moderate depth and frequent raw-signal inputs, while training practices and model inspection remained inconsistently reported. Only a minority of studies inspected trained models, most often through weights or activations.

  • 3.5 Deep learning methodology: 41% of papers used CNNs, compared with about 14% using RNNs and 13% using AEs; CNN+RNN combinations appeared in 7%.
  • 3.5 Deep learning methodology: Since 2015, CNNs became the most common architecture, replacing DBNs and FC networks that dominated through 2014.CNNs were also used more often with raw EEG, whereas deep belief networks typically used frequency-domain features and GANs used raw data.
  • 3.5 Deep learning methodology: 128 studies used architectures with at most 10 layers, while 16 did not report architecture depth.In one mental-workload study, seven-layer models outperformed two-, four-, and ten-layer models across accuracy, precision, F-measure, and G-mean.
  • 3.5 Deep learning methodology: Almost half the papers used no pre-training, 25% used pre-training, and 26% did not explicitly describe the training strategy.Standard training jointly learns representations and downstream models, whereas some studies trained feature extractors separately using RBMs, DBNs, or AEs.
  • 3.5 Deep learning methodology: More than half the studies employed at least one regularization method, often combining techniques such as dropout, L1-, and L2-regularization.Regularization was defined as constraining neural-network functions to control complexity and improve generalization on unseen data.
  • 3.5 Deep learning methodology: 47% of studies did not report the optimizer, while 30% used Adam and 17% used stochastic gradient descent; almost 80% omitted hyperparameter-search information.Adam use increased from 31.9% in 2017 to 52.6% in 2018.
  • 3.6 Inspection of trained models: 27% of studies reported inspecting their trained models, with two studies focusing specifically on model inspection in DL and EEG.
  • 3.6 Inspection of trained models: The most frequent inspections analyzed first-layer weights or model activations, while backpropagation maps and EEG-adapted class activation maps exposed influential inputs or neuron sensitivities.Activation inspection visualized network responses to known inputs, and extended CAMs highlighted image regions contributing most to decisions.

3.7 Reporting of results · 3.8 Reproducibility

The review finds that comparisons of deep learning with traditional EEG methods vary substantially in baselines, metrics, validation, and statistical testing, yet reports a median accuracy gain of 5.4%. Reproducibility remains limited because public data and code are uncommon.

  • 3.7.1 Type of baseline: 67.9% of studies included at least one traditional EEG processing pipeline as a baseline, though some also compared against deep learning approaches.Traditional pipelines typically combine feature extraction with shallow or classical machine learning models.
  • 3.7.2 Performance metrics: Accuracy, sensitivity, f1-score, ROC AUC, and precision were common metrics, while balanced accuracy, f1-score, and ROC AUC were noted as robust to class imbalance.Non-classification studies also used purpose-specific measures such as inception score and Fréchet inception distance.
  • 3.7.3 Validation procedure: Studies used varied cross-validation procedures, and some mentioned cross-validation without providing details about the method.The review grouped random train/validation/test permutations and incompletely described splits under ‘Train-Valid-Test’.
  • 3.7.4 Subject handling: 61% of studies focused only on inter-subject classification, compared with 25% focusing only on intra-subject classification and 8.3% on both.Only 22 of 108 inter-subject studies used Leave-N-Subjects-Out or Leave-One-Subjects-Out procedures, which can better reflect unseen-subject scenarios.
  • 3.7.4 Subject handling: 38% accuracy on unseen subjects versus 75% on seen subjects illustrates the difficulty of classifying EEG from subjects absent during training.The review also reports that cross-subject models often performed worse than within-subject models.
  • 3.7.5 Statistical testing: 19.9% of studies used statistical tests to compare model performance with baseline models, most often Wilcoxon signed-rank tests followed by ANOVAs.The review emphasizes that baseline, metric, validation, and reporting differences make comparisons imprecise.
  • 3.7.6 Comparison of results: 5.4% was the median gain in accuracy for deep learning over traditional baselines across 102 studies/tasks, with an interquartile range of 9.4%.Only four values were negative, while the largest improvement was 76.7% in an RSVP classification task.
  • 3.8 Reproducibility: 7% of studies could easily be reproduced using both the same data and code, while 3% were only partially reproducible.Across all studies, 54% used public data and 19% made source code available online; reproducing results may additionally require network weights.

4 Discussion

The review finds that deep learning can match or exceed traditional EEG approaches, while emphasizing shallow architectures, transfer learning, transparent data reporting, and improved reproducibility and interpretability.

  • Performance and data: Only six of 156 reviewed papers reported lower performance for deep learning than traditional benchmarks.Shallower architectures were often preferred when data were limited.
  • Generalization: Transfer learning may help overcome EEG’s subject variability, although intra-subject models still work best when data are limited.Hybrid models can be trained on pooled subjects and then fine-tuned.
  • Reproducibility: Many studies inadequately described their EEG data, making their findings hard or impossible to evaluate and compare.Recommended reporting includes subjects, sessions, and window length, alongside learning curves.
  • Inputs and preprocessing: Raw EEG increasingly serves as model input, with recent studies sometimes outperforming feature-based baselines without extensive preprocessing.CNN studies most often used raw EEG, and raw-data models improved seizure-classification performance over classical baselines in some cases.
  • Architectures: Almost half of the reviewed papers used CNNs, while most architectures contained fewer than five layers.The review concludes that shallower networks are currently necessary for EEG data compared with deep computer-vision models.
  • Reproducibility and interpretation: 21% of papers used hyperparameter search, and 27% examined model interpretation.The review calls for clearer search-method reporting and sustained efforts to understand the patterns underlying model decisions.

5 Conclusion

This review analyzed 156 DL-EEG studies to identify field trends, data and methodological variability, reported results, and reproducibility challenges. It concludes with recommendations and openly shared resources to support fairer, more reproducible research.

  • Review scope: 156 studies published between January 2010 and July 2018 were analyzed across data, EEG processing, DL methodology, reported results, and reproducibility.The review examined studies applying deep learning to EEG data.
  • Major trends: 1) DL was mainly used to classify EEG in brain-computer interfacing, sleep, epilepsy, cognitive, and affective monitoring.These applications reflect major domains identified in the review’s trend analysis.
  • Major trends: 1 to over 16,000 subjects, 62 to 9,750,000 examples, and two to 4,800,000 minutes of EEG recording show substantial variability in dataset size.The reported means were 223 subjects, 251,532 examples, and 62,602 minutes; medians were 13, 14,000, and 360, respectively.
  • Recommendations: 6 recommendations target reproducibility and fair comparison, including clear architecture and data descriptions, existing datasets, strong baselines, recording sharing, and code sharing.The review identifies code sharing as the best way to allow others to continue the work.
  • Open resources: The review’s data-items table, statistics-and-figures reproduction code, and full paper summaries are openly available online at dl-eeg.com.These resources are intended to help the DL-EEG community maintain an up-to-date list of published work.

Appendices · Appendix A List of acronyms

Appendix A defines the acronyms used throughout the paper, covering EEG-related modalities, analytical methods, neural-network architectures, evaluation metrics, and optimization procedures.

  • Appendix A List of acronyms: BCI denotes brain-computer interface, and RSVP denotes rapid serial visual presentation.
  • Appendix A List of acronyms: CCA, ICA, and PSD denote canonical correlation analysis, independent component analysis, and power spectral density.
  • Appendix A List of acronyms: CNN, DBN, DNN, RNN, and LSTM denote convolutional, deep belief, deep neural, recurrent, and long short-term neural-network concepts.
  • Appendix A List of acronyms: EEG means electroencephalography, while ECoG and MEG mean electrocorticography and magnetoencephalography, respectively.
  • Appendix A List of acronyms: ERP, SNR, and STFT denote event-related potential, signal-to-noise ratio, and short-time Fourier transform.
  • Appendix A List of acronyms: GAN, RBM, and SDAE denote generative adversarial network, restricted Boltzmann machine, and stacked denoising autoencoder.
  • Appendix A List of acronyms: NLP denotes natural language processing, SGD denotes stochastic gradient descent, and SVM denotes support vector machine.
  • Appendix A List of acronyms: ROC AUC denotes area under the receiver operating curve.

Appendix B Checklist of items to include in a DL-EEG study

Appendix B presents a checklist of information that DL-EEG papers should clearly report to make their published results readily reproducible.

  • Purpose: The checklist is intended to improve reproducibility of published DL-EEG results.The authors frame the checklist as guidance for what future studies should report.
  • Reporting location: Future DL-EEG studies should state all checklist items clearly in the main text or supplementary materials.The passage allows the required information to appear at one point or another across these materials.

Data

The data section emphasizes reporting subject demographics, electrode montage, example shape, augmentation details, and dataset splits. These details support clear characterization and implementation of EEG studies.

  • Report the number of subjects and relevant demographic data.
  • Specify the electrode montage, including references, channel count, and channel locations.
  • Describe one example’s shape, data augmentation technique, and the numbers of examples in training, validation, and test sets.

EEG processing

EEG processing considers whether studies use temporal filtering, spatial filtering, artifact handling, and resampling. These preprocessing choices are reported individually.

  • EEG processing: Temporal filtering is recorded when applicable.
  • EEG processing: Spatial filtering is recorded when applicable.
  • EEG processing: Artifact handling techniques are recorded when applicable.
  • EEG processing: Resampling is recorded when applicable.

Neural network architecture

The neural network architecture section emphasizes reporting the number of layers and learnable parameters, optionally using a diagram or table to represent the architecture.

  • Neural network architecture: Report the number of layers in the neural network architecture.A diagram or table may help represent the architecture.
  • Neural network architecture: Report the number of learnable parameters.

Training hyperparameters

The training hyperparameters section identifies core optimization and training settings that should be reported. It also emphasizes documenting stopping, regularization, search procedures, and exact hyperparameter values, including random seeds.

  • Training reports should specify parameter initialization, loss function, batch size, and number of epochs.
  • Optimization details should include the stopping criterion, regularization method, optimization algorithm, learning-rate schedule, and optimizer parameters.
  • Reported results should include all hyperparameter values, including the random seed, together with the hyperparameter search method.

Performance and model comparison

Performance and model comparison should report evaluation metrics, validation schemes, and baseline-model descriptions. These details support clear comparison across EEG studies.

  • Performance metrics such as F1-score and accuracy should be reported.
  • Studies should specify whether validation was intra-subject, inter-subject, leave-one-subject-out, or k-fold cross-validation.
  • Baseline models should be thoroughly described or linked to published references.
Loading 1901.05498v2…