fNIRSAtlas

About

fNIRSAtlas is the largest open benchmark for fNIRS classification: 23 classification tasks from 19 public data collections, with 375 unique participants and 29,438 epochs. A single automated pipeline takes every task from raw data to evaluation, under both cross-subject and within-subject protocols.

Six task suites (motor, mental arithmetic, motor imagery, n-back, emotion, miscellaneous) feed a standardized data format and any ML model, evaluated with cross-subject and within-subject splits.

Features

Automatic download, preprocessing, and classification
Nested cross-validation for evaluation and hyperparameter search
Baselines: LDA, ANN, CNN, LSTM, EEGNet, DCN, MultiRocketHydra, fNIRS-T, fNIRSNet23, fNIRSNet24
Run your own PyTorch or scikit-learn models
Statistical comparison between models
Processed data readable with no heavy dependencies

License

Code: MIT. Datasets keep their original licenses, listed below. Check them before use, some restrict derivatives or commercial use.

Code

Source, documentation, and examples are on GitHub.

Install

pip install "fnirsatlas[all]"

Run a suite

From a repository checkout. Downloads, processes, and classifies the data.

python scripts/run_all.py --suite ma

Load processed data

from fnirsatlas.io import fNIRSAtlasDataset

ds = fNIRSAtlasDataset("bnci")
X, y, groups, info = ds.as_tuple()

Datasets

The pipeline fetches data from its original sources. Raw data takes about 60 GB, processed data about 800 MB.

Dataset	Suite	Subjects	Epochs	Channels	Conditions	Hz	Device	License
Buccino2016	Motor	15	1500	68	4	10.4	NIRx NIRScout 8-16	CC BY 4.0
Bak2019	Motor	30	2250	40	3	13.3	Shimadzu LIGHTNIRS	CC BY 4.0
Ortega2020	Motor	14	356	48	3	12.5	NIRx NIRScout	MIT
Luke2021	Motor	5	450	40	3	7.8	NIRx NIRScout 16x24	Other (open)
Cockx2023	Motor	24	1844	48	16	50	Artinis Brite24	CC BY-SA 4.0
Guevara2024	Motor	20	395	44	2	50	Artinis Brite MKII	CC BY 4.0
Luke2024b	Motor	10	1081	128	4	8.9	NIRx NIRScout	CC BY 4.0
Gao2024	Motor	12	1080	200	2	8.7	NIRx NIRSport2 16x16	CC0
Bnci	Mental arithmetic	8	348	104	2	10	Hitachi ETG-4000	CC BY-ND 4.0
Shin2016 (ma)	Mental arithmetic	29	1740	72	2	10	NIRx NIRScout	GPL-3.0
Shin2018b	Mental arithmetic	12	720	18	2	10	NIRx NIRScout	CC BY 4.0
Shin2019 (ma)	Mental arithmetic	18	1620	32	3	13.3	Shimadzu LIGHTNIRS	CC BY 4.0
Shin2016 (mi)	Motor imagery	29	1740	72	2	10	NIRx NIRScout	GPL-3.0
Shin2019 (mi)	Motor imagery	18	1620	32	3	13.3	Shimadzu LIGHTNIRS	CC BY 4.0
Yi2025	Motor imagery	18	5735	48	8	7.8	NIRx NIRScout	CC BY 4.0
Herff2014	N-back	10	300	16	3	25	Artinis Oxymon Mark III	On request
Shin2018 (n-back)	N-back	26	702	72	3	10	NIRx NIRScout	CC BY 4.0
Huang2021	N-back	68	1088	4	4	5.2	ISS Imagent	CC BY 4.0
Nemo (empe)	Emotion	31	1203	48	4	50	Artinis Brite24	CC BY-NC 4.0
Nemo (afim)	Emotion	31	718	48	4	50	Artinis Brite24	CC BY-NC 4.0
Shin2018 (wg)	Misc	26	1560	72	2	10	NIRx NIRScout	CC BY 4.0
Shader2021	Misc	8	368	104	3	3.9	NIRx NIRScout	CC BY 4.0
Luke2024a	Misc	17	1020	50	3	5.2	NIRx NIRScoutX	CC BY 4.0
Total		375	29438

Subjects total counts unique participants. Epochs and conditions cover all recorded conditions. Herff2014: access granted automatically after signing a form.

Citation

@inproceedings{makela2026fnirsatlas,
  title     = {fNIRSAtlas: A Large-Scale Benchmark for Functional Near-Infrared Spectroscopy Classification},
  author    = {M{\"a}kel{\"a}, Kalle and Ruotsalo, Tuukka},
  booktitle = {Advances in Neural Information Processing Systems, Datasets and Benchmarks Track},
  year      = {2026}
}