A task EEG dataset with preprocessed signals and features for ADHD research in children and adolescents

Abstract

This dataset provides preprocessed electroencephalography (EEG) signals and extracted EEG features for research on attention-deficit/hyperactivity disorder (ADHD) in children and adolescents. Derived from version 2 of the BALLADEER dataset, it contains 669 eligible task recordings from 156 participants recruited into ADHD and comparison groups. The resource combines continuous EEG derivatives, aligned analysis windows, and 115 feature definitions or templates across seven feature families, represented in 107,478,560 primary feature-table records, together with source-linked quality and processing records. Feature definitions expand across channels, channel pairs and other indexed positions; the stored rows represent repeated measurements and structured outputs, not independent participants or observations. Participants and task conditions The source inventory contains 158 participant metadata records and 683 available EEG recordings. Following eligibility and quality checks, 669 recordings from 156 participants are represented in the preprocessed signals and feature outputs. The five task conditions are: BAL01: CogniFit cognitive assessment. BAL02: Attention Robots. BAL03: Slackline, level 1. BAL04: Slackline, level 6. BAL05: Slackline, level 11. These are task-based recordings; the resource contains no dedicated resting-state condition. Recruitment group and assessment-derived status are preserved as separate metadata fields. Participants recruited into the comparison group should not automatically be interpreted as clinically confirmed healthy controls. What the dataset contains Preprocessed EEG signals: 669 continuous recording derivatives in MNE-FIF format, stored at 128 Hz, with channel information, boundary annotations and recording-level processing provenance. Aligned analysis windows: gap-aware windows of 2, 4, 8 and 16 seconds, with identifiers and sample boundaries supporting comparisons across feature families without crossing recording gaps. EEG feature tables: seven documented families comprising 115 feature definitions or templates. These expand over channels, channel pairs and other indexed positions; they do not constitute a single table of 115 predictor columns. Quality and eligibility records: source inventories, inclusion and exclusion decisions, and quality summaries before and after preprocessing. Documentation and provenance: feature catalogues, calculation states, processing parameters, source identifiers, manifests and SHA-256 checksum inventories. Seven EEG feature families Temporal and statistical features. Conventional spectral features. Aperiodic spectral features. Complexity and fractal features. Functional connectivity features. Graph and network features. Time-frequency decomposition features. Feature availability depends on window duration and estimator requirements. For example, aperiodic spectral outputs are supported at 8 and 16 seconds. Calculation-state records distinguish unavailable or undefined values from valid numerical results. Null values should not automatically be replaced with zero. Research and educational reuse The combination of preprocessed EEG signals, documented feature tables and quality annotations supports scientific research, methodological studies and practical teaching. Potential applications include: EEG research: investigate task-related EEG characteristics and associations with participant metadata, compare feature representations on matched intervals, and examine how signal quality relates to feature values. Feature extraction studies: use the preprocessed signals to implement alternative feature calculations and compare their outputs with the supplied features, while accounting for window duration, channel configuration and estimator requirements. Feature selection studies: use the existing feature tables to compare filter, wrapper and embedded selection methods. Research questions may address feature redundancy, the stability of selected subsets across participant splits, and the relationship between subset size and predictive performance. Teaching and practical exercises: use selected tasks and feature families in courses on biomedical signal processing, neuroscience, data science and machine learning. Exercises can cover EEG feature extraction, feature selection, missing-value interpretation, data integration and participant-level cross-validation. The supplied feature tables allow reuse without repeating the documented preprocessing and feature-extraction steps. Researchers and students can begin with a single task and feature family, then expand their analysis using the shared identifiers and catalogues. Task conditions, acquisition devices and channel configurations differ. A common sampling rate does not establish device equivalence, and pooled analyses cannot automatically separate task effects from device effects. Preprocessing does not guarantee artifact-free EEG. Recruitment group and assessment-derived status must remain distinct when defining research questions or prediction targets. Repeated windows, channels and recordings from the same participant are dependent observations. Within each evaluation split, all data from a participant must remain in a single training, validation or test partition. Imputation, scaling, feature selection and model tuning must be learned from the training data, using inner validation where required, without using the held-out test data. This is particularly important when comparing feature-selection methods. Files and getting started The package contains 48 files: 41 ZIP archives comprising one metadata and quality archive, five task-specific preprocessed EEG archives, and 35 task-by-feature-family archives; plus seven supporting files covering documentation, manifests, checksums, the publication audit, version information and licensing. Read README.md and consult ZENODO_FILE_INVENTORY.csv to identify the required archives. Verify downloaded files against SHA256SUMS. Consult the feature and window catalogues before selecting durations, expanding indexed values or joining tables. Retain participant, recording, window and channel identifiers, together with the relevant calculation states and quality records. Source data and attribution This is a secondary resource derived from the BALLADEER multimodal dataset, version 2 . Original raw recordings and non-EEG modalities remain in that source deposit and are not duplicated here. Shared pseudonymous identifiers support linkage to compatible source records; event-level synchronization requires verification of the available timing information. The dataset is distributed under the Creative Commons Attribution 4.0 International licence . Please cite this derivative dataset and the original BALLADEER dataset when using these resources.