Skip to content
cEEGrid Guide

Mikkelsen et al., 2019

sleep/wake machine learning

为什么重要

  • 路线定位:around-the-ear EEG / sleep/wake machine learning。
  • 任务或证据:Sleep/wake random forest vs PSG/actigraphy
  • 自研用途:Use PSG labels, LOSO validation, and derived cEEGrid channels.

Evidence Matrix Summary

FieldValue
Route / hardwareBilateral cEEGrid sleep ML
Task / evidence baseSleep/wake random forest vs PSG/actigraphy
Main findingAutomatic cEEGrid beat actigraphy/manual cEEGrid and approached PSG-derived systems.
Key limitationHealthy small cohort; WASO/REM latency weak.
Use for our systemUse PSG labels, LOSO validation, and derived cEEGrid channels.

PDF Download

下载 PDF

静态路径:/papers/16-mikkelsen-2019.pdf

Detailed Reading Card

基本信息

  • 年份/出处: 2019, Journal of Sleep Research, 28:e12786.
  • DOI: 10.1111/jsr.12786.
  • 路线: around-the-ear cEEGrid sleep/wake machine learning。
  • 本地文件: library/pdfs_by_category/04_sleep_hearing_real_world/16_2019_mikkelsen_et_al_machine_learning_derived_sleep_wake_staging_from_around_the_ear_electroencephalo.pdf

研究问题

  • 目标是评估基于 cEEGrid 的自动睡眠分期是否优于人工 cEEGrid 评分和 actigraphy,并与人工 PSG scoring 比较(Abstract; Intro; PDF pp. 1-2)。
  • 关键前提: cEEGrid 振幅特征不同于标准 PSG,可能不适合直接人工视觉评分,因此需要机器学习(Abstract; Intro; PDF p. 1)。

硬件系统

  • PSG: SomnoHD,128 Hz,F3/F4/O1/O2/C3/C4 对侧 mastoid reference,另有 ECG/EOG/chin EMG(Recording setup; PDF p. 3)。
  • cEEGrid: 每耳 10 electrodes;R4a/R4b common ground/reference;SMARTING wireless amplifier,250 Hz;Sony Z1 smartphone bedside recording(Recording setup; PDF p. 3)。
  • 阻抗策略: >50 kOhm 的 cEEGrid 电极从分析中丢弃,不再调整,以保证 cEEGrid 安装 <10 min(Recording setup; PDF p. 3)。

电极点位 / 布局

  • 候选 derivations 包括 intra-C pairs、L-R、top-bottom、front-back;最终采用 FB(L)、FB(R)、L-R 三个标准导联(Derivation selection; Fig. 2-3; PDF pp. 3, 6-7)。
  • 选择依据是 correlation index: 与 PSG band power 相关并按电极可靠性加权(Sec. 3.2; PDF p. 3)。

实验设计

  • 与 Sterr 2018 同一类 extended sleep opportunity: 20 人招募,最终 15 人,6 male,age 35 +/- 14.3,range 18-63;12 h in-bed protocol,单晚、无 adaptation night(Participants; PDF p. 2)。
  • Actiwatch 同步记录,佩戴于非优势手;cEEGrid 与 PSG 同时采集约 12 h(Recording setup; PDF p. 3)。
  • 数据量: 自动评分使用 18,920 epochs,平均每人 10.5 h,range 3.1-12.0 h(Results; PDF p. 5)。

信号处理流程

  • EEGLAB v13.5.6b;PSG 与 cEEGrid resample 到 256 Hz;0.5-100 Hz band-pass,50 Hz notch;用 movement artifacts / slow waves 对齐 PSG 与 cEEGrid(Preprocessing; PDF p. 3; Appendix 1)。
  • 特征: 33 features x 3 derivations,包括 time-domain、EMG/EOG proxies、frequency-domain、sleep-event proxies;cEEGrid 自动评分使用 99 features(Features; Table 1; PDF pp. 4-5)。
  • 分类器: random forest,100 decision trees,Matlab 2016b fit ensemble Bag algorithm;leave-one-subject-out validation;ground truth 默认人工 PSG scoring(Random forest classifier; PDF p. 5)。
  • 后处理: sleep onset 需 5 min 连续 sleep,wake up 需前方 5 min sleep,class probabilities 用 5-epoch moving average 平滑(Hypnogram post-processing; PDF p. 5)。

结果

  • Table 4: 人工 PSG 中 Wake 34.8%、pooled sleep 65.2%;自动 cEEGrid Wake 38.4%、pooled sleep 61.6%,比 actiwatch Wake 25.5% 更接近 PSG(PDF p. 6)。
  • 二分类 sleep-wake 和五分类 sleep staging 中,自动 cEEGrid 的 kappa 显著优于 cEEGrid* 和 actiwatch;与 automatic PSG、cEEGrid+EOG、PSG DOMINO 差异不显著(Table 5; PDF p. 8)。
  • Sleep statistics: cEEGrid 对 total sleep 的 ICC 0.933,sleep efficiency 0.750,WASO 0.690,SOL 0.771;REM latency 表现差,ICC -0.101(Table 7; PDF p. 9)。
  • 特征数分析显示最佳约 20-30 features,max kappa 0.58;使用全部 99 features kappa 约 0.53,过拟合代价较小(Fig. 7; PDF p. 10)。

局限

  • 15 名 healthy good sleepers 样本仍小;需更大 cohort、临床 cohort、老年健康人群验证后才能替代 PSG(Discussion; PDF p. 10)。
  • WASO 和 REM latency 仍不可靠;作者认为短暂 arousal 与 proper wake EEG 特征差异可能导致 WASO 问题(Discussion; PDF p. 10)。
  • manual cEEGrid scoring 不能替代 simultaneous PSG labels 作为算法 ground truth;非标准 montage 需要 PSG 同步训练标签(Discussion; PDF p. 9)。

对自研的启发

  • cEEGrid 睡眠产品不宜只依赖人工 AASM 评分;算法应使用同步 PSG 标签训练,并明确验证 leave-one-subject-out 或外部测试集。
  • 导联压缩可以从高密度 cEEGrid 中选 FB(L)、FB(R)、L-R 这类稳健 derivations,而不是随机选单电极。
  • Actigraphy 便利但特异性有限;around-ear EEG 可在可用性与准确性之间提供更好折中。

下一批建议

  • 第四批优先整理 17-21:Reali 2021、Knierim 2022、Holtze 2022、Knierim 2023、Van Den Broucke 2023。
  • 重点转向 OpenBCI/cEEGrid adapter、amplifier benchmarking、continuous-speech AAD 与 wireless hardware。

Metadata

FieldValue
IDp16_mikkelsen_2019_sleep_wake
TitleMachine-learning-derived sleep-wake staging from around-the-ear electroencephalogram outperforms manual scoring and actigraphy
Year2019
Category04_sleep_hearing_real_world
Routearound-the-ear EEG
Stagesleep/wake machine learning
Statusprocessed
Source integrityok
Pages13
OCR statusnot_needed

Evidence Groups

GroupHitsPages
hardware12p. 1, p. 2, p. 3
electrode_layout8p. 1, p. 2, p. 3, p. 6, p. 7, p. 11, p. 12
experiment12p. 1, p. 2, p. 3
signal_processing12p. 1, p. 2, p. 3
results12p. 1, p. 2
limitations12p. 2, p. 3, p. 4, p. 5

Local Evidence Sources

  • Source PDF path: US-pdf/Journal of Sleep Research - 2018 - Mikkelsen - Machine‐learning‐derived sleep wake staging from around‐the‐ear.pdf
  • Public PDF path: /papers/16-mikkelsen-2019.pdf
  • Categorized PDF path: library/pdfs_by_category/04_sleep_hearing_real_world/16_2019_mikkelsen_et_al_machine_learning_derived_sleep_wake_staging_from_around_the_ear_electroencephalo.pdf
  • Extracted text path: library/texts/04_sleep_hearing_real_world/16_2019_mikkelsen_et_al_machine_learning_derived_sleep_wake_staging_from_around_the_ear_electroencephalo.txt
  • Detailed card source: library/DETAILED_PAPER_CARDS_BATCH_3.md
  • Page-level evidence index: library/EVIDENCE_INDEX.md

Close Reading Checklist

  • Verify exact figures, tables, page numbers, and statistics against the local PDF before formal citation.
  • Keep missing parameters as Not reported unless the PDF or supplementary material confirms them.
  • Mark any cross-paper synthesis as interpretation rather than a single-paper claim.