High-Fidelity Synthetic Dataset for EEG-Based Eye State Detection (Open vs Closed Eyes)This dataset contains the optimal high-fidelity synthetic version of the CD_52 dataset, derived from the EEG Eye State Dataset (UCI Machine Learning Repository). The original data comprises 14 EEG channel measurements from an Emotiv EEG Neuroheadset, recorded continuously for 117 seconds (14,980 samples). The channels follow the International 10-20 electrode placement system: AF3, F7, F3, FC5, T7, P7, O1, O2, P8, T8, FC6, F4, F8, AF4. The target variable is eyeDetection, indicating whether the subject's eyes were open (0) or closed (1), as determined by simultaneous video recording. The data is strictly chronological, representing a continuous time series of brain activity.Generation and Selection Methodology:The synthetic data was generated using the rigorous 4-stage methodology detailed in the accompanying research article, "Methodology for generating high fidelity synthetic data to improve mortality prediction for patients in intensive care units". Following an initial class balancing stage, a large-scale synthetic generation process was executed using 10 different methods.The final dataset was selected by the TabDSFidelity index, a novel metric designed to balance statistical fidelity and predictive utility. This dataset corresponds to the SMOTE RSB Refinado* algorithm, which achieved the highest TabDSFidelity score (7.961) among all pure synthetic candidates, demonstrating superior preservation of the original data's statistical properties and predictive patterns. The evaluation was performed against the original reference dataset to ensure the highest fidelity.File Details:Format: CSV (Comma-Separated Values)Number of Records: 7,000 synthetic instancesNumber of Attributes: 15 (14 EEG channels + 1 target variable)Target Variable: eyeDetection (1 = Eyes closed, 0 = Eyes open)Positive Class Ratio: The dataset is balanced to mitigate the original class imbalance (Ratio: 1.000).Intended Use:This dataset is intended for research and development of machine learning models for brain-computer interfaces (BCI), EEG signal processing, and neuroscience applications. It is suitable for training and evaluating binary classification algorithms where accurate detection of eye state from EEG signals is critical for applications such as driver fatigue monitoring, sleep studies, cognitive state assessment, and assistive technologies for individuals with motor disabilities. The temporal nature of the data also makes it suitable for sequence models (LSTM, GRU, Transformers). The synthetic nature of the data allows for safe sharing and benchmarking without exposing potentially sensitive biological data.Performance Summary:Models trained on the original data achieved a weighted F1-score of 0.64, AUC-ROC of 0.90, and accuracy of 0.65. The original model showed moderate performance on both classes, with precision of 0.62 and recall of 0.54 for the eyes closed class (class 1). Models trained on the optimal synthetic dataset achieved a weighted F1-score of 0.99 (+0.35 improvement), AUC-ROC of 1.00 (+0.10 improvement), and accuracy of 0.99 (+0.34 improvement). The synthetic data dramatically improved both classes, with precision increasing from 0.62 to 0.98 (+0.36) and recall from 0.54 to 1.00 (+0.46) for the eyes closed class. This demonstrates the exceptional effectiveness of the methodology in capturing complex EEG signal patterns and achieving near-perfect classification performance, which is critical for practical BCI applications.EEG Channel Information:The 14 EEG channels correspond to the International 10-20 system:Frontal: AF3, AF4, F7, F3, F4, F8, FC5, FC6Temporal: T7, T8Parietal: P7, P8Occipital: O1, O2The occipital channels (O1, O2) are typically the most discriminative for eye state detection due to the alpha rhythm (8-13 Hz) increase when eyes are closed.Data Preprocessing Note:The original dataset (14,980 EEG samples) was obtained from the UCI Machine Learning Repository. The ARFF file was parsed manually to extract the @data section, with column names assigned according to the official documentation. EEG channel values were converted to float using pd.to_numeric(..., errors='coerce'), and the target variable eyeDetection was converted to integer. Rows with null values from conversion were removed to ensure data integrity. A stratified subsample of 1,000 samples was created for efficient experimentation while preserving the original class distribution (~55% eyes open / ~45% eyes closed).Temporal Nature Note:The data is a continuous time series of 117 seconds duration. Samples are in chronological order, with the first value measured at the beginning of the recording. This temporal structure makes the dataset suitable for sequence modeling approaches. When using the data for classification, it is recommended to use time-series validation (e.g., time-series split) rather than random cross-validation to prevent data leakage between training and test sets.Privacy and Ethics:The data is entirely synthetic and does not contain any real EEG recordings that could be traced to specific individuals. The original dataset was fully anonymized with no personally identifiable information. A post-generation privacy audit was conducted, which included evaluating the risk of re-identification via Distance to Closest Record (DCR) and a Membership Inference Attack (MIA). The results showed a low risk of membership inference (MIA score: 0.437, below random guessing), and zero exact copies were generated, making this dataset suitable for broader research sharing, while acknowledging that perfect privacy cannot be guaranteed and use should be governed by ethical guidelines for synthetic biological data.Related Work:This dataset is part of a larger project on synthetic data generation for biomedical signal processing and brain-computer interface applications. For more details on the methodology and additional datasets, please visit the project's companion website: https://mdiazbastida.github.io/codigomdb/investigaciones.html