Neural Noise

Simultaneous EEG and fNIRS for semantic decoding of imagined animals and tools

2 min read

  • A multimodal EEG+fNIRS dataset supports semantic decoding of imagined categories (animals vs tools). 1
  • The dataset is published in Nature Scientific Data and targets non-invasive decoding pipelines and reproducible benchmarks. 1
  • Direct BCI relevance: supports building and benchmarking non-invasive semantic decoding pipelines. 1 1

Weekly enrichment (2026-07-20)

  • The source is a data descriptor by Rybář, Poli & Daly (University of Essex) in Scientific Data (Nature portfolio, 2025; DOI 10.1038/s41597-025-04967-0), motivated by building a semantic BCI that communicates whole concepts directly, bypassing the slow character-by-character spelling of current BCIs.2
  • The paradigm discriminates two semantic categories — animals versus tools — across four mental tasks: a silent naming task plus three sensory-based imagery tasks (visualizing the object, imagining its sounds, and imagining touching it).2
  • Dataset 1 contains simultaneous EEG and fNIRS from 12 participants across all four tasks; Dataset 2 is a simplified follow-up with EEG-only recordings from 7 additional participants on the auditory imagery task alone.2
  • fNIRS was acquired with a NIRx NIRScoutXP continuous-wave system (8 sources/emitters and 4 detectors with low-profile optodes) using two montages (frontal and temporal); EEG and fNIRS optodes shared an integrated NIRx NIRScap layout.23
  • The EEG side comprises 64 EEG channels plus 11 misc and 1 trigger channel; the release is BIDS-formatted and hosted on OpenNeuro as ds004514 (v1.1.2, 137 files, ~24.2 GB).34
  • The authors frame EEG and fNIRS as complementary and portable/low-cost alternatives to fMRI: EEG contributes millisecond temporal resolution while fNIRS adds ~2 cm-depth hemodynamic spatial information, improving ecological validity for real-world semantic BCIs.2
  • A prior fNIRS-only study by the same group (Rybář et al., J. Neural Eng. 18(4):046035, 2021) reported mean two-class accuracies of 76.2% (silent naming), 80.9% (visual imagery), 72.8% (auditory imagery), and 70.4% (tactile imagery) using logistic regression on GLM-extracted hemodynamic responses, and showed classifiers generalize across tasks.5
  • Methodological caveat: the same authors later warned (Rybář et al., Sci. Rep. 14:28003, 2024) that using data from cue-presentation windows grossly overestimates semantic-BCI performance — a benchmarking pitfall this open dataset lets others control for.3
  • BCI implication: the dataset supports building and reproducibly benchmarking non-invasive, multimodal semantic-decoding pipelines aimed at intuitive communication for people who cannot speak.2

Footnotes

  1. https://news.google.com/rss/articles/CBMiX0FVX3lxTE5PS0plNmFrNDlfdl95RE5nRDdWMmRIOVRLSl9wMHNYSklWNjk2OS1lWkVhaGhKWmY2dDV1YVVZenkyUEd6YVpYYk9CZkpzVkNwNExFYzlKb0tEdC1kOVRr?oc=5 ↩ ↩2 ↩3 ↩4

  2. https://www.nature.com/articles/s41597-025-04967-0 ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  3. https://nemar.org/dataexplorer/detail?dataset_id=ds004514 ↩ ↩2 ↩3

  4. https://openneuro.org/datasets/ds004514/versions/1.1.2 ↩

  5. https://iopscience.iop.org/article/10.1088/1741-2552/abf2e5 ↩

Graph View