ear-eeg-fm-benchmark-data
收藏资源简介:
该数据集是“Ear-EEG FM Benchmark”的预处理版本,以HDF5格式提供,旨在支持耳-EEG(耳周脑电图)与头皮EEG(脑电图)的基准测试和比较研究。数据集包含两个主要任务:听觉注意检测(AAD)和睡眠阶段分类。对于AAD任务,提供二元分类(左 vs 右)数据,包括头皮EEG变体(saad-eeg.h5)和耳-EEG变体(saad-ear-eeg.h5),分别包含27个和14个通道,采样频率为1000 Hz,试验次数分别为13,500次(头皮)和未明确次数(耳,但形状一致)。对于睡眠分类任务,提供5类睡眠阶段数据,包括头皮PSG变体(eesm23-eeg.h5)和耳-EEG变体(eesm23-ear-eeg.h5),分别包含8个和4个通道,采样频率为250 Hz,试验次数分别为未明确次数(头皮)和56,967次(耳)。每个HDF5文件遵循统一架构:data字段为形状(N, C, T)的float32数组,表示以微伏(µV)为单位的EEG信号;labels字段为int64类别标签;subject和session字段标识受试者和会话;trial_id字段为试验标识符;ch_names字段列出标准10-20系统的通道名称;attrs字段包含元数据如采样频率、类别数、类别名称、单位和预处理配置。数据集适用于脑机接口、听觉神经科学、睡眠研究和机器学习模型开发,特别侧重于耳-EEG的便携式应用与传统头皮EEG的性能对比。
This dataset is a preprocessed version of the Ear-EEG FM Benchmark provided in HDF5 format, designed to support benchmarking and comparative studies between ear-EEG (ear-electroencephalography) and scalp EEG (electroencephalography). It includes two main tasks: auditory attention detection (AAD) and sleep stage classification. For the AAD task, binary classification (left vs. right) data is provided, including scalp EEG variants (saad-eeg.h5) and ear-EEG variants (saad-ear-eeg.h5), with 27 and 14 channels respectively, a sampling frequency of 1000 Hz, and trial counts of 13,500 (scalp) and an unspecified number (ear, but consistent in shape). For the sleep classification task, 5-class sleep stage data is provided, including scalp PSG variants (eesm23-eeg.h5) and ear-EEG variants (eesm23-ear-eeg.h5), with 8 and 4 channels respectively, a sampling frequency of 250 Hz, and trial counts of an unspecified number (scalp) and 56,967 (ear). Each HDF5 file follows a unified structure: the data field is a float32 array of shape (N, C, T) representing EEG signals in microvolts (µV); the labels field contains int64 category labels; subject and session fields identify the subject and session; trial_id field serves as a trial identifier; ch_names field lists channel names from the standard 10-20 system; and attrs field includes metadata such as sampling frequency, number of classes, class names, units, and preprocessing configurations. The dataset is suitable for brain-computer interfaces, auditory neuroscience, sleep research, and machine learning model development, with a particular focus on comparing the performance of portable ear-EEG applications with traditional scalp EEG.
数据集概述
该数据集为 Ear-EEG FM Benchmark 预处理后的 HDF5 文件,源自 Ear-EEG-FM-Benchmark 项目。所有 .h5 文件遵循统一数据模式,包含 EEG 数据、标签、受试者信息、会话信息、试验 ID、通道名称及采样率等属性。
数据模式
每个 .h5 文件包含以下字段:
- data:形状为
(N, C, T),数据类型float32,单位 µV - labels:形状为
(N,),数据类型int64,取值范围 0 到 n_class-1 - subject:形状为
(N,),数据类型object,格式sub-NN - session:形状为
(N,),数据类型object,格式ses-NN - trial_id:形状为
(N,),数据类型int64 - ch_names:形状为
(C,),数据类型object,为标准 10-20 电极名称 - attrs:包含
sfreq、n_class、class_names、unit、preprocess_config_json等属性
文件命名规则
- 文件名含
-eeg.h5:头皮 EEG 变体 - 文件名含
-ear-eeg.h5:耳部 EEG 变体
数据集文件详情
| 文件 | 任务 | 试验数 | 数据形状 | 采样率 | 来源 |
|---|---|---|---|---|---|
saad-eeg.h5 |
二元 AAD(左 vs 右) | 13,500 | (N, 27, 4000) | 1000 Hz | https://zenodo.org/records/16536441 |
saad-ear-eeg.h5 |
二元 AAD(cEEGrid 耳周) | - | (N, 14, 4000) | 1000 Hz | https://zenodo.org/records/16536441 |
eesm23-eeg.h5 |
5 类睡眠(头皮 PSG) | - | (N, 8, 1000) | 250 Hz | https://openneuro.org/datasets/ds005178 |
eesm23-ear-eeg.h5 |
5 类睡眠(cEEGrid 耳部 EEG) | 56,967 | (N, 4, 1000) | 250 Hz | https://openneuro.org/datasets/ds005178 |
使用方式
该数据集配合基准测试流程使用,具体命令为:
bash git clone git@github.com:zhikaili1150/Ear-EEG-FM-Benchmark.git cd Ear-EEG-FM-Benchmark mkdir -p data huggingface-cli download <this repo> --repo-type dataset --local-dir data DATASET=saad-eeg bash scripts/run_benchmark.sh




