README.md

---
license: cc-by-4.0
tags:
  - eeg
  - auditory-attention-decoding
  - event-related-potential
  - in-ear-eeg
  - single-word-classification
---

# EventAAD-Processed

Minimally preprocessed HDF5 export of **EventAAD Paradigm 3**
(`speech~speech~focus`) for single-word auditory-attention decoding. The
experiment presents two simultaneous natural-speech streams and asks the
listener to attend to one of them. This release contains 24 subjects, 20
experimental trials per subject, and one fixed EEG epoch per selected word.

The two files contain the same 30,739 examples in exactly the same row order,
once with scalp EEG and once with in-ear EEG. They can therefore be evaluated
separately or paired by row.

## Task and label construction

This export implements the **single-word Task B** setting from Nguyen et al.:
detect an attended target word against words from the unattended speech
stream.

Each source `events.tsv` word is anchored at its temporal midpoint
(`onset + duration / 2`). Labels are encoded as:

- `1`, `attended`: the word is in the attended stream **and** its event value
  equals that trial's attended target category (AT).
- `0`, `unattended`: the word is in the unattended stream **and** its event
  value differs from that trial's attended target category.

This yields 2,611 attended and 28,128 unattended epochs, reproducing the class
counts reported in the paper's Table I. Twenty-four otherwise eligible
unattended epochs are omitted because the complete [-0.2, +1.0] s window would
cross a recording boundary.

**Important label-definition note:** the paper describes the negative class
in prose as all events in the unattended stream, which could be interpreted as
including unattended words whose category matches the attended target (UT).
That literal rule produces 28,680 complete negative epochs in the released
EventAAD files, not the paper's 28,128. The files here use the table-count-
aligned reconstruction above: AT versus the retained unattended non-matching
words. This choice is explicit so experiments do not silently conflate the two
possible Task B interpretations.

The source word itself and its variable duration determine the epoch center,
but **not** the model input length: every example is a fixed 1.201 s signal
window from -0.2 s through +1.0 s, inclusive, around the word midpoint.

## Minimal benchmark preprocessing

The same deliberately small signal-processing front end used by the EEGFM
benchmark is applied to each full continuous recording before epoching:

- zero-phase 0.1-100 Hz band-pass;
- 50 Hz notch filter (the recordings were collected in Denmark);
- retain the native 1000 Hz sampling rate;
- no ICA, re-reference, baseline correction, amplitude rejection, resampling,
  or data augmentation.

Filtering the continuous signal before slicing avoids filter-edge artifacts at
each short epoch boundary. Scalp names `T7-T3`, `T8-T4`, `P7-T5`, and `P8-T6`
are normalized to `T7`, `T8`, `P7`, and `P8`. The scalp file uses 30 canonical
EEG channels and excludes mastoids M1/M2, EOG, and `cRef`; the in-ear file
retains all 12 left/right ear electrodes.

### How this differs from the single-word paper

Nguyen et al.'s task-specific scalp-EEG pipeline used 32 scalp channels,
downsampled 1000 Hz recordings to 256 Hz, applied a zero-phase FIR 0.5-40 Hz
filter, removed ocular artifacts with ICA, formed [-0.2, +1.0] s epochs around
the word midpoint, and rejected epochs exceeding 200 microvolts peak-to-peak.
Its training procedure also used random ERP averaging and synthetic ERP
augmentation.

Those later steps are intentionally not baked into this release. Keeping the
native rate and a common broad passband makes this dataset comparable with
other EEGFM benchmark datasets and leaves model-specific time/frequency
transforms to the downstream pipeline. Avoiding ICA and amplitude rejection
also prevents task-specific cleaning choices and modality-dependent row loss,
so scalp and ear examples remain paired. Augmentation belongs inside the
training split and should never be applied to validation or test examples.

## Files

| file | channels | examples (N) | sfreq | epoch |
|---|---:|---:|---:|---:|
| `eventaad-task-b-scalp-eeg.h5` | 30 scalp EEG | 30,739 | 1000 Hz | 1.201 s (1201 samples) |
| `eventaad-task-b-in-ear-eeg.h5` | 12 in-ear EEG | 30,739 | 1000 Hz | 1.201 s (1201 samples) |

Both files contain 28,128 label-0 and 2,611 label-1 rows. There is one
Paradigm-3 recording/session per subject and 480 experimental trials in total
(24 subjects x 20 trials). `trial_id` identifies the actual parent experimental
trial (`0` through `19`), `event_id` identifies the source `events.tsv` word
row, and `sample_id` identifies the processed HDF5 row. `split_group_id` equals
`trial_id` because a trial is the indivisible split unit.

### Leakage-safe protocol grouping

Word epochs from the same continuous experiment are not independent: nearby
words can share most of their source EEG samples. Split by
`(recording_id, split_group_id)`, never by individual HDF5 row. The benchmark's
default multi-subject and within-subject temporal split keeps trials 0-13 in
train, 14-16 in validation, and 17-19 in test. This gives 21,427 / 4,368 /
4,944 word epochs respectively, with both classes in every subject-level
partition and zero shared source samples across partitions. LOSO instead keeps
the held-out subject intact.

`window_start_sample` and `window_stop_sample` provide half-open source-
recording bounds for independently auditing overlap. If custom fractions put a
boundary between adjacent experimental trials whose word windows overlap,
those trials must be merged into one indivisible connected group.

## HDF5 schema (v0.4)

```
/data          (N, C, T)   float32   signal, microvolts
/durations     (N,)        int64     valid samples per epoch (1201 here)
/nan_fraction  (N,)        float32   fraction of non-finite samples in the epoch
/labels        (N,)        int64     index into attrs['class_names']
/sample_id     (N,)        int64     processed HDF5 row identity
/subject       (N,)        str       'sub-001' ...
/session       (N,)        str       empty (no separate BIDS session entity)
/task          (N,)        str       'speech~speech~focus'
/acquisition   (N,)        str       empty (both channel subsets share one recording)
/run           (N,)        str       empty (no separate BIDS run entity)
/recording_id  (N,)        str       continuous source-recording identity
/trial_id      (N,)        int64     experimental trial, 0...19
/event_id      (N,)        int64     row index in the source events.tsv
/split_group_id (N,)       int64     equals trial_id
/window_start_sample (N,)  int64     start in source recording
/window_stop_sample  (N,)  int64     exclusive stop in source recording
/ch_names      (C,)        str       channel names

attrs: sfreq, n_class, class_names (['unattended','attended']), unit ('uV'),
       eegfm_version, source_bids_path, bids_dataset_name,
       preprocess_config_json, split_group_kind ('trial'),
       trial_id_kind ('experimental_trial'),
       event_id_kind ('source_events_tsv_row'), sample_id_kind,
       window_reference, created_at
```

`preprocess_config_json` records the exact channel list, label rule, filter
settings, epoch window, and the preprocessing operations intentionally left
unset.

## Data-quality note

This is a minimal-processing benchmark export, not a manually curated clinical
dataset. Some source in-ear channels contain extreme-amplitude segments. They
are retained rather than silently removing epochs or breaking scalp/ear row
alignment. Downstream experiments should fit any robust scaling or quality-
control thresholds on the training partition only and report those choices.

## Download

```bash
hf download Zachary1150/EventAAD-Processed --repo-type dataset \
  --local-dir EventAAD-Processed
```

## Sources

- Original data: [EventAAD on Zenodo](https://zenodo.org/records/8082867),
  released under CC BY 4.0.
- Dataset paper: Geirnaert et al., *Cognitive component of auditory attention
  to natural speech events*, Frontiers in Human Neuroscience (2024),
  [doi:10.3389/fnhum.2024.1460139](https://doi.org/10.3389/fnhum.2024.1460139).
- Single-word task: Nguyen et al., *Single-word Auditory Attention Decoding
  Using Deep Learning Model* (2024),
  [arXiv:2410.19793](https://arxiv.org/abs/2410.19793).

When using these processed files, please cite the original EventAAD dataset and
the relevant task paper in addition to this repository.