iMINDBench — iEEG Multi-Institution Neural Decoding Benchmark

iMINDBench provides a multi-institution, multi-task, preprocessor-structured benchmark to evaluate iEEG Foundation Models' ability to benefit decoding regardless of recording institution or task, under comparable preprocessing tracks.

3 datasets and institutions 15 aligned decoding tasks 2+ preprocessing tracks

Benchmark Design

Three naturalistic datasets, explicit preprocessing tracks, and multiple decoding tasks spanning language, vision, and auditory tasks.

Overview of the three iMINDBench datasets, neural preprocessing routes, decoding tasks, temporal splits, scaling domains, and subject subsets. — Datasets (Neuroprobe / Brain Treebank, Bang! You’re Dead, Pippi) can be easily downloaded from the torch_brain package. Recordings are harmonized and flow through standardized Multi-STFT, waveform, or custom preprocessing routes. Models are evaluated under common temporal splits, data-scale regimes, and Main / Challenge subject subsets.

Dataset Diversity

Scale, task support, and electrode coverage vary substantially across the benchmark as expected from multi-institution datasets.

Summary of training samples, decodable subject-sessions, electrode counts, task coverage, and electrode locations across iMINDBench datasets. — Our leaderboard features aggregated scores, but also decomposed comparisons by dataset, task, and preprocessing track.

BibTeX

@misc{chau2026imindbench,
  title={{iMINDBench}: {iEEG} Multi-Institution Neural Decoding Benchmark},
  author={Geeling Chau and Saba Hashemi and Yonghyeon Gwon and Eshani Patel and Jan DeWitt and Christopher Wang and Andrii Zahorodnii and Sabera J Talukder and Danny Dongyeop Han and Chun Kee Chung and Maryam M Shanechi and Yisong Yue},
  year={2026},
  eprint={2609.18104},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={http://arxiv.org/abs/2609.18104},
}