Note
Go to the end to download the full example code or to run this example in your browser via Binder.
Find a small cohort with the EEGDash API#
Search real catalogue records before downloading EEG. We select the same three Nakanishi2015 SSVEP participants used by the following tutorials. This page needs internet for metadata only; signal files are not opened.
Before you start#
Install EEGDash with its dependencies and run this file in Python or a
notebook with network access. Familiarity with a pandas DataFrame is useful;
no MNE, PyTorch or model-training knowledge is needed. The result is a bounded
cohort query that you can pass to EEGDashDataset in tutorial 01.
import pandas as pd
from eegdash import EEGDash
1. Query an explicit cohort#
A catalogue row describes one recording file. subject, session and
run are identifiers, so keep values such as "0" as strings. The
$in expression accepts any of the three named participants; the other
fields must all match. This bounds the later signal acquisition before we
open a recording.
limit=10 caps recording rows, not participants. Inspect the returned
identifiers before downloading; do not silently take the first three rows.
client = EEGDash()
query = {
"dataset": "nm000118",
"subject": {"$in": ["1", "2", "3"]},
"session": "0",
"run": "0",
"task": "ssvep",
}
records = pd.DataFrame(client.find(query, limit=10))
records[["dataset", "subject", "session", "run", "task"]]
2. Inspect the signal metadata before committing to downloads#
Read sampling_frequency as samples per second, nchans as recorded
channels, and ntimes as time samples per channel. Thus ntimes divided
by sampling_frequency estimates a file’s signal duration. These fields
help plan memory and window lengths, but they do not establish signal quality.
A participant may contribute multiple recordings; duration estimates the download workload, not signal quality.
fields = ["subject", "sampling_frequency", "nchans", "ntimes"]
records["duration_minutes"] = records["ntimes"] / records["sampling_frequency"] / 60
records[fields + ["duration_minutes"]]
3. Broaden discovery without treating a limited sample as a census#
An unfiltered sample is useful for discovering dataset and task names. It cannot estimate how common a task is across the whole catalogue: the first 20 rows need not be randomly sampled or balanced across datasets. Choose a specific dataset from discovery, then query its relevant recordings before summarizing a cohort.
sample = pd.DataFrame(client.find({}, limit=20))
sample[["dataset", "subject", "task"]]
# Change the dataset/task query to explore another paradigm, then inspect
# its real participant and event metadata before choosing a decoding target.
