A dataset for Emotion Recognition using EEG and MUsical Stimuli

# EREMUS: A Dataset for Emotion Recognition Using EEG and Musical Stimuli

## Description

EREMUS is a dataset designed for the study of emotion recognition using electroencephalography (EEG) data collected during musical stimuli exposure. The dataset includes EEG recordings from 34 young subjects in a controlled laboratory setting. Each subject participated in 16 trials, with a duration of approximately 90 seconds each. Eight trials featured songs from the subject's personal music playlist, while the other eight were randomly selected from the preferences of other participants. Following each trial, subjects self-assessed their emotional responses using the Geneva Emotion Wheel, which categorizes emotions into 20 families arranged in a circular format based on valence and dominance.

### Data Collection

EEG data was recorded using a 32-channel EPOC Flex EEG system with saline sensors, at a sampling rate of 128 Hz. Electrode placements adhered to the international 10-20 system.

### Training Dataset

The training set consists of 294 trials from 26 subjects, with each subject contributing approximately 12 labeled trials. Each trial includes the following metadata:

- **session_type**: Indicates whether the trial is from a *personal* or *other* session.- **subject_id**: Identifier for the subject.- **spotify_track_id**: Spotify identifier for the song.- **song_title**: Title of the song played.- **song_author**: List of song authors.- **emotion**: Self-assessed emotion based on the Geneva Emotion Wheel.- **label**: Emotion label in a Valence-Dominance space.- **id**: Trial identifier.

### Test Dataset

The test dataset is divided into two parts: held-out trials and held-out subjects, both released without emotion labels or subject identifiers.

- **Held-out Trials**: Comprises 104 trials from the 26 subjects included in the training dataset.- **Held-out Subjects**: Contains 122 trials from 8 subjects not present in the training dataset.

Additionally, the held-out trials dataset includes 44 extra trials for the subject identification task, which may contain or omit stimulation information.

### Preprocessing

The dataset is available in two versions:

1. **Raw EEG Data**: Unprocessed data.2. **Pruned Data**: Preprocessed using EEGLab, which includes a FIR filter applied between 0.5 and 40 Hz and artifact removal via Independent Component Analysis (ICA).

### Files Structure

The dataset is organized in a hierarchical structure according to data type (raw or pruned) and split type (training, held-out trials, held-out subjects):

```dataset├── splits_subject_identification.json├── splits_emotion_recognition.json├── README.md├── raw│   ├── train│   │   ├── 1135903657_eeg.fif│   │   └── ...│   ├── test_trial│   └── test_subject└── pruned    ├── train    │   ├── 1135903657_eeg.fif    │   └── ...    ├── test_trial    └── test_subject```

The IDs for the test sets are specified in two JSON files:- `splits_subject_identification.json` for Task 1.- `splits_emotion_recognition.json` for Task 2.