EEGEmotions-27

Description

It was collected and released by Phuong Huy-Tung, Im Eun-Tack, Oh Myeong-Seok, and Gim Gwang-Yong, and published as a data resource in IEEE Access (vol. 13, 2025) to give the affective-computing community an EEG corpus with higher emotional resolution than existing public sets such as DEAP or SEED. The recordings use a 14-channel Emotiv EPOC X headset, and the per-trial CSV column headers (EEG.AF3, EEG.F7, …) confirm the data were exported through EmotivPRO.

Citation

• Paper DOI: 10.1109/ACCESS.2025.3620677
• Dataset DOI: TODO(confirm)
• Reference: Phuong, H.-T., Im, E.-T., Oh, M.-S., & Gim, G.-Y. (2025). EEGEmotions-27: A Large-Scale EEG Dataset Annotated With 27 Fine-Grained Emotion Labels. IEEE Access 13.

Catalog citation block

• Phuong, H.-T., Im, E.-T., Oh, M.-S., & Gim, G.-Y. (2025). EEGEmotions-27: A Large-Scale EEG Dataset Annotated With 27 Fine-Grained Emotion Labels. IEEE Access, 13. DOI: 10.1109/ACCESS.2025.3620677. (Article CC BY 4.0; dataset CC BY-NC 4.0)
• Cowen, A. S., & Keltner, D. (2017). Self-report captures 27 distinct categories of emotion bridged by continuous gradients. PNAS, 114(38), E7900-E7909. DOI: 10.1073/pnas.1702247114. (Source of the 27-emotion taxonomy.)

Download and access

Field	Value
Official source	TODO(confirm)
Access class	open
Size	TODO(confirm)
Format	csv, txt
BIDS	false
Mirrors	none recorded

Source & access

Primary publication: IEEE Access, vol. 13 (2025), DOI 10.1109/ACCESS.2025.3620677 (IEEE Xplore document 11202184). The IEEE article-of-record is open access (CC BY 4.0), but the dataset is released under CC BY-NC 4.0 per the paper’s data-availability statement and the repository README. A copy of the paper is also mirrored on ResearchGate.

Data structure

• Format: flat per-trial CSV + TXT pairs (no BIDS, no DataLad/git-annex). The filename pattern is Subject_Trial.csv / Subject_Trial.txt (e.g. 10_1.0.csv, 10_1.0.txt).
• CSV files carry a header row and a leading sample-index column, then 14 named EEG channels (EEG.AF3 … EEG.AF4) — the standard EmotivPRO CSV export.
• TXT files contain the same 14-channel signal as bare tab-separated values with no header (one fewer row than the matching CSV, since the CSV header line is absent).

License

• Name: CC-BY-NC-4.0
• Note (verbatim from the directory): dataset CC BY-NC 4.0 per the paper’s data-availability statement; the IEEE article-of-record is CC BY 4.0; download location TODO(confirm)
• Snippets on this site: no — the site ships no snippet or precomputed product from this dataset; lessons link the official access page.

Acquisition

Field	Value
Device	14-channel wireless saline-electrode headset (Emotiv EPOC X)
Device class	consumer
Channels	14
Sampling rate (Hz)	256
Online filters	device hardware default (≈ 0.16–43 Hz Sinc anti-alias; 50/60 Hz notches); not documented beyond the default
Reference	TODO(confirm)
Mains frequency (Hz)	TODO(confirm)
Paradigms	emotion-video
Participants	88
Population	88 participants; demographics (age, sex) undocumented
Clinical groups	none
Sessions	1
Durations	27 trials per subject (~25–70 s each)

Acquisition

EmotivPRO was configured at 256 Hz per channel (verbatim in Phuong et al., 2025 and the repo README); the headset oversamples internally (2048 Hz) and applies a built-in Sinc anti-alias filter, giving an effective hardware bandwidth of roughly 0.16—43 Hz, with hardware digital notch filtering at 50 Hz and 60 Hz. Recording length is per-trial rather than fixed: sampled files run from roughly 6,000 to 18,000+ rows (~25 s to ~70 s at 256 Hz), consistent with variable-length video clips. Online reference/montage details beyond the EmotivPRO default are not documented.

Channels

14 EEG channels in the standard Emotiv EPOC X 10-20 montage (column order as exported by EmotivPRO):

AF3, F7, F3, FC5, T7, P7, O1, O2, P8, T8, FC6, F4, F8, AF4

Participants

88 participants are recorded in the published dataset (Phuong et al., 2025). No clinical group is described; this is a healthy-population affective-elicitation study.

Tasks / conditions

Participants viewed emotionally evocative short video clips and the resulting EEG was annotated against the 27 discrete emotion categories of Cowen & Keltner’s (2017) semantic-space theory of emotion — e.g. amusement, awe, fear, sadness, joy, disgust — which is a finer-grained labelling scheme than the valence/arousal or positive/neutral/negative schemes common to other EEG emotion sets (Phuong et al., 2025). The dataset’s accompanying deep-learning baseline reported an average classification accuracy of 62.24% across the 27 categories on 1,168 held-out samples (Phuong et al., 2025).

Used on this site

L7.5

Loader

Fetches only the stated subset. Recorded loader: TODO(confirm).

TODO(confirm) — the catalog records no loader for this dataset.

Caveats worth knowing

• Nominal 256 Hz output but ≈ 43 Hz hardware bandwidth and built-in 50/60 Hz notches — the Nyquist frequency is not the usable bandwidth.

Notable caveats

• Effective bandwidth is ~43 Hz. Although sampled at 256 Hz (128 Hz Nyquist), the EPOC X Sinc filter rolls off above ~43 Hz, so there is no meaningful neural content in the top of the band.
• Subjects 26 and 48 are missing, and a handful of trials are absent, so do not assume a complete 88 × 27 grid.
• Per-subject demographics and the exact public download mirror are not documented in the sources consulted; the IEEE Access data-availability statement is the authoritative pointer.

Related datasets

• MultiPENG (EEG stream)
• ArEEG_Words
• EEG for natural-image recognition (VEP)
• PhySF, Physiological Sense of Flow
• ERP CORE (Compendium of Open Resources and Experiments)

# Loader: TODO(confirm) — the catalog records no loader for this dataset.
# Download a small subset from the official source: TODO(confirm)