科研 Skill

癫痫发作检测评估

neuro-seizure-detection-eval

Evaluates EEG or iEEG seizure detection and prediction algorithms the way clinicians and challenges score them. Use when reporting sensitivity, false alarms per hour, or latency.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

Seizure detection evaluation

Sample- or window-level accuracy is meaningless for seizures: they are rare, so a detector that never fires scores above 99%. Clinical usefulness is about catching each seizure quickly without too many false alarms.

1. Split by patient, and by time within patient

  • Patient-independent claims: leave-one-patient-out.
  • Patient-specific claims: train on earlier seizures, test on later ones (never random windows; windows around one seizure are near-duplicates).
  • Prediction/forecasting: also hold out the interictal data in time order, and keep a gap around each seizure as the dataset protocol defines.

2. Score events, not windows

Convert detector output to events (threshold, smoothing, minimum duration, merge gaps), then compare with reference seizures:

  • Sensitivity: fraction of reference seizures overlapped by a detection.
  • False alarms per hour (FA/h, or per 24 h): detections that overlap no seizure, divided by recording hours.
  • Latency: detection onset minus reference onset, for detected seizures.
  • Precision and F1 on events as secondary numbers.

scripts/event_scoring.py computes all of these from two CSVs:

python scripts/event_scoring.py --ref ref.csv --det det.csv --hours 72.5 [--tolerance 0]

Each CSV has onset_s,offset_s (optionally record for multi-file data).

Use any-overlap scoring unless the user's benchmark defines otherwise (TUH and challenge papers sometimes use time-aligned or epoch-based variants; report which).

3. Choose the operating point honestly

Pick the threshold on training data, then report test sensitivity and FA/h at that threshold. A full sensitivity vs FA/h curve is useful, but the headline must be one fixed operating point.

4. Report

Patients, recording hours, seizures per patient, split scheme, threshold choice, event sensitivity, FA/h, median latency, and per-patient results (averages hide patients where it fails).

示例提示词

帮我按临床标准评估这个癫痫检测算法的灵敏度和误报率。

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

癫痫发作检测评估能做什么?

Evaluates EEG or iEEG seizure detection and prediction algorithms the way clinicians and challenges score them. Use when reporting sensitivity, false alarms per hour, or latency。

如何安装癫痫发作检测评估?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「癫痫发作检测评估」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。

癫痫发作检测评估需要哪些权限?

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。

癫痫发作检测评估经过官方审核吗?

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。