Deep Learning of EEG Signals for Brain Function and Injury: A Systematic Review

This dataset is the result of a systematic review focused on Deep Learning (DL) applications for Electroencephalography (EEG) in the context of Brain Function and Injury (BFI), including consciousness assessment, brain injury, and comatose states. The data were collected following the PRISMA 2020 protocol and PROSPERO registration (CRD42023489566), covering studies published between 2018 and 2023 in major scientific databases such as ACM Digital Library, Elsevier, Google Scholar, IEEE, PubMed, and Scopus.

Each record in the dataset represents a study that met strict eligibility criteria, including the use of human EEG data, DL techniques, and the provision of details on datasets, preprocessing methods, feature extraction, classifier architectures, and performance metrics. Studies combining EEG with other modalities (e.g., fMRI, ECG) or lacking methodological transparency were excluded.

The dataset contains rich metadata organized into more than 80 fields, separated by the "#" delimiter. These fields cover:

Context and task (e.g., classification, prediction, clinical context)

Dataset characteristics (public availability, sample size, duration, arrangement, channels, frequency bands)

Preprocessing techniques (noise removal, dimensionality reduction, normalization, data augmentation)

Feature extraction methods and technical details

Deep Learning architectures (CNN, LSTM, GAT, GAN, Transformers) and hyperparameters (e.g., layers, activation functions, optimizers, learning rates, batch sizes, epochs)

Evaluation strategies (validation type, dataset division, cross-validation folds, parameter optimization)

Performance metrics (accuracy, AUC, sensitivity, specificity, precision, F1-score, MAE, Wasserstein distance, reconstruction error, spectral entropy difference)

Explainable AI (XAI) considerations and comparisons with other techniques or datasets.

This dataset is intended to support reproducibility, transparency, and meta-analysis in EEG-based DL research. It enables the identification of trends, methodological gaps, and opportunities for improvement, such as advanced preprocessing strategies, diverse DL architectures, robust validation procedures, and the integration of XAI for interpretability in clinical contexts.

Researchers can use this resource for benchmarking, comparative analysis of hyperparameters, exploration of performance variability across contexts, and the development of new DL models tailored to the specific challenges of EEG in BFI applications.