Datasets:

---

NeuroTec

/

SWEC_iEEG_Dataset

like 3

Follow

NeuroTec 4

Thanks for releasing this dataset!

While onboarding this dataset, I came across two separate issues that I want to bring to your attention.

Heads up, I asked Claude to draft an outline of my discovery but I edited and checked all the content below myself first.

1. _total.h5 virtual dataset has broken pointers for 50/68 participants

The dataset card recommends reading each participant's signal through
<ID>_total.h5. For ID01-ID18 this works as intended, however, for all the remaining
participants (ID19-ID68), the _total.h5 files point to the wrong references.
They're all off by 18, for instance:

• ID19_total.h5's sources point at ID01_part_1.h5

Since those referenced files don't exist inside the affected participant's own
folder, and HDF5 silently fills the data zeros instead of raising an error.

To reproduce:

import h5py

with h5py.File("ID19_total.h5", "r") as f:
    dset = f["data/ieeg"]
    for vsrc in dset.virtual_sources():
        print(vsrc.file_name)   # -> "ID01_part_1.h5" (should be ID19_part_1.h5)
    print([n.decode() for n in f["info/files"][:]])  # -> correct names

info/files correctly names each participant's own part files so it's usable
as a workaround, but the _total.h5 itself needs to be regenerated with the correct
source paths.

2. 18 of the 68 participants are exact duplicates

The following 18 folders reproduce one of ID01-ID18 exactly -- identical
channel count, sampling rate, data/ieeg shape, and (checked to the reported
floating-point precision) identical seizure onset/offset annotations:

I confirmed this isn't just matching metadata: I checked three of the pairs
(ID04/ID24, ID09/ID30, ID10/ID31) and downloaded both sides'
_part_N.h5 payloads and compared their data/ieeg arrays
directly (numpy.array_equal, part-by-part). They were identical

Btw, I think that the info/checksums didn't catch it because there might be
some metadata that is distinct (filename embedded) that would lead to the checksum
being different.

This means the headline "68 participants / 9,328 hours / 704 ictal events"
double-counts the original 18-participant SWEC cohort; de-duplicated, the
release is 50 participants / 6,672 hours / 460 ictal events.

Anyone doing a participant-level train/test split, or reporting per-participant statistics,
would want to know about this to avoid leakage between participants that are actually the same recording.

Happy to share the verification scripts if it's helpful useful or submit a PR.
I have generated the corrected _total.h5 if you would like them.
I have created a patch locally for this in our workflow, but it'd be great to fix here.