Shared reference/ground strategy: 两个系统都有 Fp1 shared electrode 以便后处理合并 reference system;ground shared at CPz(EEG data acquisition; PDF p. 4)。
电极点位 / 布局
Around-ear array 为 C-shaped around/behind-ear layout;左侧底部靠近 cheek 的一个 around-ear electrode 被牺牲,用作 Fp1 shared electrode(Around-ear EEG; PDF p. 4)。
In-ear electrodes 遵循 Kidmose naming convention:ear canal 两个 ExI/ExE,tragus 一个 ExT,concha 三个 ExA/ExB/ExC;x 表示左右耳,具体位置因耳廓解剖差异略变(In-ear EEG; PDF p. 5)。
Scalp EEG 包含 10-20 位置和 EOG;Fp1 在 scalp 与 ear-EEG 系统间共享,CPz 共享 ground(Fig. 2/Table 1; PDF pp. 4-5)。
实验设计
被试: 15 名 self-reported normal hearing adults,7 women、8 men,19-31 岁,mean 24.7,SD 3.5,native Danish speakers;Aarhus University IRB approval 2024-0673174(Participants; PDF p. 3)。
任务: selective auditory attention to one of two competing speech streams,左右注意方向 counterbalanced;每名 6 trials x 10 min,总计 60 min AAD data(Protocol; PDF p. 3)。
刺激: 两路 Danish educational talks,由 6 名 male Danish professors 讲授;每 trial 前有 5 s target prelisten;speech 通过 integrated insert earphones 呈现,并用 HRTF 模拟 ±60 deg azimuth;音量 normalized to -33 LUFS(Protocol/stimuli; PDF pp. 3-4)。
视觉条件: 半数 trials 显示 attended speaker video,半数显示 fixation cross;video/no-video 与左右方向跨被试平衡(Protocol; PDF pp. 3-4)。
触发同步: audio trigger channel 和 video stream photosensor triggers;实验末还采 5 min broadband chirp ASSR,但本论文不分析 ASSR(Protocol; PDF p. 4)。
信号处理流程
AAD algorithm: classic linear stimulus reconstruction;用 EEG backward decoder 重构 attended speech envelope,并比较与两路 competing speech features 的 Pearson correlation(Methods; PDF pp. 5-6)。
Decoder: EEG time lags 0-400 ms post-stimulus;L2 regularization,lambda 使用 Ledoit-Wolf analytical estimator(Methods; PDF p. 6)。
EEG preprocessing: 各 setup 分开处理;1-9 Hz zero-phase 4th-order Butterworth bandpass;按 trigger 分割 10 min trials(Preprocessing; PDF p. 6)。
Downsample/normalization: downsample to 20 Hz,per-trial Frobenius norm normalization;默认每个 setup 内 common average reference,around-ear/in-ear CAR 时排除 shared Fp1(Preprocessing; PDF pp. 6-7)。
Speech feature: gammatone filterbank 19 bands from 50 Hz to 5 kHz,powerlaw compression exponent 0.6,sum envelope,1-9 Hz filter,downsample to 20 Hz,per-trial zero mean/unit variance(Speech feature extraction; PDF p. 7)。
Evaluation: participant-specific leave-one-trial-out CV;participant-independent leave-one-participant-out CV;decision windows 1, 5, 10, 30, 60, 120, 300, 600 s;主比较使用 60 s(Evaluation; PDF p. 7)。
结果
主性能阶梯: 60 s decision window average AAD accuracies 为 scalp EEG 83.44%,around-ear EEG 67.22%,in-ear EEG 61.11%;三者差异显著,显示 performance vs wearability trade-off(Abstract/Conclusion; PDF pp. 1, 15)。
个体显著性: 60 s 下 scalp 15/15 participants significant,around-ear 14/15 significant,in-ear 9/15 significant;Benjamini-Hochberg correction 后 around-ear 13/15,in-ear 6/15(Results; PDF pp. 8-9)。
Speed limitation: median MESD 为 scalp 35.4 s,around-ear 2.66 min,in-ear 7.35 min;即使 scalp linear SR 对 very short windows 也不足,ear-based 更不适合当前算法下的快速 gain switching(Results; PDF pp. 9-10)。
Participant-independent decoding: scalp 有典型下降,60 s 下降 7.67 percentage points;in-ear 跨被试泛化失败并低于显著,around-ear 只下降 1.67 percentage points,作者推测标准化 C-shaped array 降低了 inter-participant variability(Results; PDF p. 10)。
Complementarity: 把 around-ear/in-ear 加到 full scalp EEG 中没有提高 AAD performance,原因是 performance gap 和 neural tracking correlation gap 太大(Complementarity; PDF pp. 10-11)。
Reference effects: same-ear average reference 使 in-ear 下降 4.1 percentage points,但 around-ear 小幅上升 1.7 points;单个 scalp reference 对 in-ear 提升 9.3 points,对 around-ear 提升 5.8 points,最佳 scalp references 在 left fronto-central area,如 FC5、C3、FC1、Cz(Influence of reference; PDF pp. 12-13)。
Sensor network: greedy forward selection 从 in-ear/around-ear 出发,添加 8 个 scalp nodes 可达到 full-scalp performance;in-ear 加 3 个 scalp nodes 接近 scalp median accuracy。in-ear + top three scalp nodes FC5/C3/T8 达 73.4%,同三 scalp electrodes 无 in-ear node 为 64.0%,差异 p = 0.0123(Sensor network; PDF pp. 13-15)。
局限
样本 15 名,且为年轻、normal-hearing、native Danish homogeneous group;对老年人和听损人群的泛化仍需验证(Conclusion; PDF p. 15)。
场景虽无电屏蔽,但仍是受控听觉任务;需要更 ecologically valid listening scenarios(Conclusion; PDF p. 15)。
每 10 min trial 只问一个 comprehension question,且 responses 未记录,无法量化 trial 内注意力波动或用任务表现筛除(Conclusion; PDF p. 15)。
使用 classic linear stimulus reconstruction;作者指出 ear-based EEG 仍缺少成功的 tailored nonlinear AAD algorithms(Conclusion; PDF p. 15)。
Scalp/around-ear 用 wet electrodes,in-ear 用 dry electrodes,因此 modality comparison 同时混入了 electrode wet/dry 和 contact quality 差异(Conclusion/Discussion; PDF pp. 9, 15)。
若目标是实时 neuro-steered hearing aid,当前 ear-based linear SR 仍不够快;若目标是长时间注意力监测,around-ear/in-ear 都已有显著可用信号。
Metadata
Field
Value
ID
p23_geirnaert_2025_direct_comparison
Title
A Direct Comparison of Simultaneously Recorded Scalp, Around-Ear, and In-Ear EEG for Neural Selective Auditory Attention Decoding to Speech
Year
2025
Category
06_recent_preprints_comparisons
Route
scalp / around-ear / in-ear EEG
Stage
direct modality comparison
Status
processed
Source integrity
ok
Pages
19
OCR status
not_needed
Evidence Groups
Group
Hits
Pages
hardware
12
p. 1, p. 2
electrode_layout
12
p. 1, p. 2, p. 3, p. 4, p. 5
experiment
12
p. 1, p. 2, p. 3
signal_processing
12
p. 1, p. 2, p. 3, p. 4, p. 6
results
12
p. 1, p. 2, p. 3
limitations
12
p. 1, p. 2, p. 4, p. 5, p. 6, p. 7
Local Evidence Sources
Source PDF path: US-pdf/A Direct Comparison of Simultaneously Recorded Scalp, Arou.pdf
Public PDF path: /papers/23-geirnaert-2025.pdf
Categorized PDF path: library/pdfs_by_category/06_recent_preprints_comparisons/23_2025_geirnaert_et_al_a_direct_comparison_of_simultaneously_recorded_scalp_around_ear_and_in_ear_eeg_f.pdf
Extracted text path: library/texts/06_recent_preprints_comparisons/23_2025_geirnaert_et_al_a_direct_comparison_of_simultaneously_recorded_scalp_around_ear_and_in_ear_eeg_f.txt