Datasets:
license: cc0-1.0
tags:
- eeg
- medical-time-series
- icml2026
- open-reproductions
task_categories:
- time-series-forecasting
ADFTD (OpenNeuro ds004504) — processed for the TS-Fingerprint reproduction
Processed arrays for the ICML 2026 reproduction of "Learning Fingerprints for Medical
Time Series with Redundancy-Constrained Information Maximization" (OpenReview
lrRBHFIgaK, arXiv
2605.00130).
Provenance
Source: OpenNeuro ds004504 — A dataset of EEG
recordings from Alzheimer's disease, Frontotemporal dementia and Healthy subjects
(Miltiadous et al., 2023), CC0. We use the derivatives/ (ASR + ICA cleaned) .set
recordings, 19 channels at 500 Hz, 88 subjects.
Processing
Reproduces data_preprocessing/ADFTD_preprocessing.ipynb of
DL4mHealth/Medformer
(the protocol the paper cites), see prep_adftd.py:
- linear-interpolation resample 500 Hz → 256 Hz
- non-overlapping 256-timestamp (1 s) windows, centre-anchored on each recording
- labels
{C: 0 healthy, F: 1 FTD, A: 2 AD} - subject-wise 60/20/20 split within each class (Medformer
ADFTDLoader, a=0.6, b=0.8) - per-sample, per-channel standardisation (Medformer
uea.normalize_batch_ts)
Contents
| File | Shape | Notes |
|---|---|---|
X_train.npy |
(40446, 256, 19) | 51 subjects |
X_val.npy |
(14658, 256, 19) | 18 subjects |
X_test.npy |
(14648, 256, 19) | 19 subjects |
y_*.npy |
(N,) int64 | 0 = CN, 1 = FTD, 2 = AD |
subj_*.npy |
(N,) int64 | subject id per window, for leakage checks |
Total 69,752 samples — this exactly matches the paper's Table 5 (#Sample = 69,752),
which is the main evidence that the preprocessing pipeline is faithful.
Please cite the original data descriptor: doi:10.3390/data8060095