科研 Skill

神经数据集陷阱审查

neuro-dataset-gotchas

Checklist for auditing neural datasets before training: rate and channel consistency, duplicate subjects, label mapping. Use before modeling on a public or lab dataset.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

Dataset audit before you train

Most surprising results on public datasets come from the data, not the model: mixed sampling rates, duplicated people, channel sets that change between files, evaluation labels that differ from training labels. Spend ten minutes on this audit before any modelling.

1. Run the audit

For each file or run, tabulate:

  • sampling rate, number and names of channels, units, reference;
  • duration, number of events per class, event code mapping;
  • subject id, session id, site, date.

Then check:

  1. Rates and channels identical across files? If not, resample and select a common channel set explicitly, and say so.
  2. One person, one id? Merged or multi-site sets can list the same person twice. Compare demographics and feature similarity.
  3. Classes balanced per subject? Pooled balance can hide subjects with one class only, which breaks per-subject CV.
  4. Labels match the documentation? Recount events against the paper or README; check for off-by-one event codes and swapped class names.
  5. Train/evaluation split defined by the dataset? Use it for comparison with published numbers; competitions often released evaluation labels separately.
  6. Data quality: flat or saturated channels, long artifact periods, discontinuities, and recordings with long wake or idle periods.

Record exclusions with reasons in a table you keep with the results.

2. Known issues to check first

These are well-known pitfalls. Confirm each against the dataset's own documentation before acting, since versions change.

DatasetCheck
BCI Competition IV 2a22 EEG + 3 EOG channels; separate training (T) and evaluation (E) sessions; evaluation labels shipped separately
PhysioNet EEG Motor Movement/Imageryseveral subjects are commonly excluded for sampling or annotation irregularities (often S088, S092, S100, S104); runs mix executed and imagined tasks, so select runs by task
Sleep-EDFlong wake periods before and after sleep inflate accuracy; trim to sleep period plus margin; older stage 3 and 4 labels map to N3
CHB-MITchannel montage changes between files of the same patient; pediatric patients; seizure annotations per file
TUH EEG corporamixed montages and sampling rates across sessions; use the corpus's train/dev/eval split and patient ids
MOABB datasetsuse MOABB's paradigms and evaluation classes so results compare with the literature

3. When merging datasets

Harmonise rate, channels, reference, filtering and units first. Keep a dataset id column and test leave-one-dataset-out, because site effects are large and a model can learn the site.

Output

An audit table (file, rate, channels, duration, events per class, issues, action) plus the final list of subjects and runs used.

示例提示词

我下载了CHB-MIT癫痫数据集,帮我检查里面常见的标注和采样陷阱。

所属插件

神经科学面向神经科学研究:EEG/MEG、fMRI、BIDS、spike sorting、连接组、脑机接口解码评估、神经刺激安全与公开神经数据检索。

安装该插件会一并获得同领域的 45 个 Skill / MCP。

同插件的其他科研 Skill

常见问题

神经数据集陷阱审查能做什么?

Checklist for auditing neural datasets before training: rate and channel consistency, duplicate subjects, label mapping. Use before modeling on a public or lab dataset。

如何安装神经数据集陷阱审查?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「神经数据集陷阱审查」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。

神经数据集陷阱审查需要哪些权限?

该科研 Skill没有声明额外权限;执行具体操作时仍由 Agent 按自身规则请求确认。

神经数据集陷阱审查经过官方审核吗?

是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。