Artifact detection and localization in single-channel mobile EEG for sleep research using deep learning and attention mechanisms

arXiv:2504.08469 · eess.SP, cs.LG · Submitted 2026-08-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Artifact detection and localization in single-channel mobile EEG for sleep research using deep learning and attention mechanisms".

Jane: The paper was written by Khrystyna Semkiv, Jia Zhang, Maria Laura Ferster and Walter Karlen from Institute of Biomedical Engineering, Ulm University and Mobile Health Systems Lab, Department of Health Sciences and Technology, ETH Zurich.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: Welcome back, everyone. We're looking at a paper that's got a mouthful of a title — "Artifact detection and localization in single-channel mobile EEG for sleep research using deep learning and attention mechanisms." Jane, let's break that down for our listeners who might've just tuned in.

Jane: Absolutely, Tom. So, when you're sleeping at home wearing one of those headbands that measures brain activity, the signal gets messy. You move, you blink, your muscles twitch — all of that creates noise that we call artifacts. This paper is about teaching a computer to spot that noise automatically.

Tom: And it's not just about spotting it — the title says "detection and localization." That's a big deal. Detection means saying "this twenty-second chunk is bad." Localization means pointing to the exact four-second window inside that chunk where the noise happened.

Jane: Right, and that matters because sleep researchers currently have to stare at hours of brainwave data and manually mark where the noise is. It's tedious, it's slow, and it's subjective. Two different experts might mark the same recording differently.

Tom: The team behind this is from Ulm University and ETH Zurich — Khrystyna Semkiv, Jia Zhang, Maria Laura Ferster, and Walter Karlen. They've been working on wearable sleep monitoring for a while, so this isn't just a theoretical exercise.

Jane: What I love about this title is the word "mobile." This isn't lab equipment with thirty electrodes glued to your head. This is a single channel, one electrode, worn at home while you sleep normally. That's a much harder problem.

Tom: Why is that harder, Jane? Isn't less data easier to handle?

Jane: Actually, it's the opposite. With many electrodes, you can compare signals and cancel out noise. With one channel, you've got nothing to compare against. It's like trying to hear one voice in a crowded room with one ear instead of two.

Tom: That's a great way to put it. And the "attention mechanisms" part of the title — that's the clever bit. The model doesn't just classify the signal; it learns to focus on the parts that matter, like a spotlight on the noisy sections.

Jane: Exactly. And that spotlight gives you the localization ability. The model can say "the artifact is right here," not just "somewhere in this chunk."

Tom: So the big question — why should our listeners care about this? Who's this paper for?

Jane: Anyone who's ever worn a sleep tracker and wondered why their sleep score seemed off. Or anyone working in sleep research who's drowning in data. And honestly, anyone building wearable health devices that need to work in the real world, not just in a lab.

Tom: And we're going to dig into how they actually built this thing and whether it works. Stay with us.

Summary: Tom: We're back with "Artifact detection and localization in single-channel mobile EEG for sleep research using deep learning and attention mechanisms." Jane, give us the quick version — what did these folks actually do?

Jane: They built a deep learning model that takes raw brainwave data from a wearable device and decides whether each twenty-second chunk contains artifacts. And if it does, it can pinpoint which four-second window inside that chunk is the noisy one.

Tom: And the data — this is real-world stuff, right? Not lab recordings?

Jane: Real-world, yes. They used data from a clinical trial where healthy older adults wore a sleep headband at home for multiple nights. We're talking about ninety-eight recordings from twenty-four participants, all sleeping in their own beds.

Tom: So this is messy, uncontrolled, real-life data. People rolling over, adjusting pillows, maybe snoring, maybe the band shifts a bit.

Jane: Exactly. And that's what makes it valuable. The artifacts in this data are the ones you actually get in practice, not the ones you'd see in a controlled lab setting.

Tom: Now, the key result — how well did it work?

Jane: Their best model, which they call CNN-CBAM, achieved an area under the ROC curve of zero point eight eight. That's a measure of how well the model separates artifacts from clean signal. Sensitivity was zero point eight one, meaning it caught eighty-one percent of the artifact chunks. Specificity was zero point eight six, meaning it correctly left alone eighty-six percent of the clean chunks.

Tom: So it's catching most of the noise and not crying wolf too often. But it's not perfect.

Jane: No, and they're honest about that. The localization part — finding the exact four-second window — was harder. Sensitivity dropped to zero point six one and specificity to zero point six three. So it's pointing in the right direction, but it's not surgical precision.

Tom: But here's what I find impressive — they compared their model against six other approaches, including some standard signal-processing methods that have been around for years. And their deep learning model beat them all.

Jane: That's the key takeaway for me. The traditional methods, like setting a threshold on signal amplitude or spectral power, are rigid. They need manual tuning and they don't adapt well to different people or different nights. The deep learning model learns what artifacts look like directly from the data.

Tom: And it does that with minimal preprocessing — no filtering, no manual feature engineering. You just feed it the raw signal and it figures it out.

Jane: Which is huge for real-world deployment. If you want to run this on a phone or in the cloud for thousands of users, you don't want to have to hand-tune parameters for each person.

Tom: So the summary is: deep learning works better than old-school methods for finding noise in home sleep recordings, and it can even point to where the noise is. Next up, we're going to talk about what makes their model special under the hood.

Improvements: Tom: Back with "Artifact detection and localization in single-channel mobile EEG for sleep research using deep learning and attention mechanisms." Jane, we've covered what they did and how well it worked. Now let's talk about what's actually new here.

Jane: The biggest improvement is the attention mechanism. They took a standard convolutional neural network and added something called a Convolutional Block Attention Module — CBAM for short. It's not a brand-new idea, but they adapted it for one-dimensional time-series data like EEG.

Tom: And what does that attention module actually do?

Jane: Think of it like a teacher grading an essay. A regular CNN reads every word with equal weight. The attention module learns which words — or in this case, which time points and which features — matter more. It highlights the important parts and ignores the rest.

Tom: And that's what gives them the localization ability. The attention map shows where the model is focusing, which turns out to be where the artifacts are.

Jane: Exactly. And here's the interesting part — they did an ablation study. That means they built several versions of the model to see which components actually help. They had a plain CNN, a CNN with LSTM layers, and then both of those with the attention module added.

Tom: And what did they find?

Jane: The attention module helped a lot. The plain CNN got an AUC of zero point seven three. Adding LSTM layers brought it up to zero point seven seven. But adding attention to the CNN jumped it to zero point eight eight. That's a huge leap.

Tom: But here's the surprise — when they combined attention with LSTM, it actually did worse than attention alone. zero point eight four instead of zero point eight eight.

Jane: That's counterintuitive, right? You'd think more complex would be better. But the LSTM adds a lot of parameters and computational cost without improving performance. The attention module already captures the temporal patterns, so the LSTM becomes redundant.

Tom: And there's a practical benefit to that. The CNN with attention runs about ten times faster than the LSTM versions — zero point zero zero six milliseconds per epoch versus zero point zero six zero. For real-time monitoring, that matters.

Jane: Another improvement they made was using SMOTE to handle the class imbalance. In their data, only about four percent of epochs contained artifacts. If you train a model on that, it'll just learn to say "clean" all the time and get ninety-six percent accuracy without actually detecting anything.

Tom: So SMOTE creates synthetic artifact examples to balance the training data.

Jane: Yes, and they did a qualitative check to make sure the synthetic samples looked like real artifacts. They compared power spectral densities and confirmed the synthetic ones preserved the general characteristics.

Tom: So the improvements are: attention mechanism for better detection and localization, a simpler architecture that runs faster, and a way to handle the imbalance problem. But what does this mean in practice? We'll get into that next.

First Page: Tom: Still with us on "Artifact detection and localization in single-channel mobile EEG for sleep research using deep learning and attention mechanisms." Jane, let's zoom in on the opening of the paper. What's the problem they're setting up?

Jane: The first page makes a really important point about why this matters. EEG is essential for studying sleep, cognition, brain-computer interfaces. But traditional EEG setups are lab-bound — high-density electrodes, expert supervision, uncomfortable for the patient.

Tom: And that's where wearable EEG comes in. Devices people can wear at home, stream data to the cloud, monitor themselves over long periods.

Jane: But here's the catch — when you move EEG out of the lab, the artifact problem gets worse. You've got electrode displacement from movement, sweat, muscle activity, eye movements. And the signal quality degrades over time as the device shifts during the night.

Tom: And the manual approach — having experts visually inspect the data — becomes impossible at scale.

Jane: Right. They mention that manual identification is time-consuming, labor-intensive, and tedious. And for large-scale deployment of wearable monitors, you can't have humans staring at every minute of data.

Tom: The paper also mentions something I hadn't thought about — the Internet of Medical Things. These devices aren't just recording data; they're connected. They can stream to the cloud in quasi-real-time. But that only works if you have automated quality control.

Jane: Exactly. If you're streaming EEG data to a cloud platform for analysis, you need to know immediately if the signal is garbage. Otherwise, you're wasting bandwidth and computing power on noise.

Tom: And there's a clinical angle too. If someone's using a wearable EEG to monitor for epilepsy or sleep disorders remotely, false readings could lead to wrong decisions.

Jane: The paper also reviews what's been tried before. There are signal-processing methods — setting thresholds on amplitude or spectral power. There are machine learning approaches. But most were designed for lab settings with high-density EEG. They don't transfer well to single-channel wearable data.

Tom: So the gap they're filling is specifically for single-channel, home-based, wearable EEG. That's a niche but growing area.

Jane: And that's why the first page sets up the problem so well. It's not just "artifacts are bad." It's "wearable EEG is the future, but it won't work without automated artifact detection, and existing methods don't cut it."

Tom: So the stage is set. Now, we've talked about the method and the results. But what does this actually mean for the world? Let's bring in the rest of the team for that.

Conclusion: Tom: Wrapping up our discussion on "Artifact detection and localization in single-channel mobile EEG for sleep research using deep learning and attention mechanisms." Jane, give us the final summary.

Jane: The team built a deep learning model that can detect artifacts in single-channel sleep EEG from wearable devices, and it can also point to where in the signal the artifacts occur. Their best model, CNN-CBAM, beat six other approaches, including traditional signal-processing methods and other machine learning models.

Tom: And the key innovation was the attention mechanism — it not only improved detection accuracy but also gave them a way to visualize and localize the artifacts.

Jane: Right. The attention maps show exactly where the model is focusing, and that correlates with where the artifacts actually are. That's a big step toward making the model interpretable, not just a black box.

Tom: So what's the real-world impact here?

Jane: For sleep researchers, this means they can process nights of data automatically instead of manually. For wearable device makers, it means they can build quality monitoring into their systems. And for patients using these devices, it means more reliable data and better-informed clinical decisions.

Tom: There are limitations, of course. The localization isn't perfect — sensitivity of zero point six one means it misses some artifacts. And the dataset is from healthy older adults, so it might not generalize to other populations.

Jane: But the framework is solid. The approach of using attention mechanisms for both detection and localization is something that could extend beyond sleep research — to any wearable biosignal monitoring.

Tom: And that's what excites me. This isn't just a paper about sleep EEG. It's a demonstration that deep learning with attention can make wearable health monitoring practical and interpretable.

Jane: Absolutely. And with the growth of remote healthcare and IoMT devices, this kind of automated quality assessment is going to become essential.

Tom: Well said, Jane. That's our take on "Artifact detection and localization in single-channel mobile EEG for sleep research using deep learning and attention mechanisms." Thanks for joining us, and we'll see you for the next paper.

Jane: Take care, everyone.

Khrystyna Semkiv, Jia Zhang, Maria Laura Ferster, Walter Karlen

Institute of Biomedical Engineering, Ulm University · Mobile Health Systems Lab, Department of Health Sciences and Technology, ETH Zurich

eess.SP, cs.LG

Submitted: 2026-08-08

License: http://creativecommons.org/licenses/by-nc-sa/4.0/

Importance score: 58/100

The gist: The paper addresses the challenge of detecting and localizing artifacts in single-channel electroencephalogram (EEG) data collected from wearable devices in uncontrolled home environments.

Terminology

Summary

The paper addresses the challenge of detecting and localizing artifacts in single-channel electroencephalogram (EEG) data collected from wearable devices in uncontrolled home environments. The authors state: Our goal was to develop and validate an automated method to detect and localize EEG artifacts, specifically to cloud-based analysis of long-term sleep data from IoMT-enabled wearables. The work focuses on designing a deep learning model to accomplish a binary detection of artifacts in raw sleep EEG obtained from remote home monitoring and integrates a machine learning attention mechanism to facilitate the easy localization, visualization, and interpretation of artifacts within a streamed EEG epoch.

The study used a single-channel EEG dataset from a clinical trial (NCT03420677) [35], in which healthy older adults wore a mobile device at home for multiple nights. The data came from 24 Caucasian participants (10 female and 14 male) with a mean (SD) age of 68.12 y (±4.72). A total of 98 recordings with a median duration of 7.9 h (range 5.5 to 9.9 h) were available. Participants wore the Mobile Health Systems Lab Sleep Band (MHSL-SleepBand v2, [9]) that sampled biosignals at 250 Hz with electrodes at the central forehead (Fpz) and both mastoid positions.

The data was split into training (58%), validation (17%), and test (25%) sets, stratified by age and number of recordings per subject. The test set contained 26 separate recordings from 6 healthy participants with a mean (SD) age of 68.33 y (±4.08), which contained artifacts in 4% of epochs.

The preprocessing pipeline involved: resampling all raw EEG signals to 128 Hz, segmenting them into 20-second non-overlapping epochs, each containing five 4-second windows, clipping the first 20 seconds of each recording due to extreme values, applying min-max scaling (mapping to [0, 1]) for each 20s-epoch separately, and applying the Synthetic Minority Oversampling Technique (SMOTE) as the last preprocessing step to the training and validation sets [36] to maintain class balance during training and parameter tuning. The authors note that Due to the heavy imbalance (approx. 5% artifacts) SMOTE was necessary.

Four deep learning models were developed with increasing complexity:

  1. Baseline two-branch CNN: It consists of two separate branches, each featuring CNN layers with small and large kernel sizes, respectively. The small kernel focuses on a short time scale to capture time-domain information while the large kernel focuses on the expanded time scale, detecting repetitive patterns and frequency information in the EEG epochs. Each branch includes five convolutional layers, each followed by batch normalization and, in some cases, by a dropout regularization mechanism and/or max-pooling operation.

  2. CNN-LSTM: The CNN-LSTM model was built based on the same model for sleep stage classification with the same adaptations and includes a single bidirectional LSTM with 128 hidden units with a shortcut connection from the CNN outputs to the LSTM output.

  3. CNN-CBAM: The baseline CNN augmented with a channel block attention module (CBAM) [37] which uses a sequential attention mechanism, applying channel and then spatial attentions one after another. The 1D-CBAM was adapted for time-series data. The channel attention captured the inter-channel relationship of the CNN feature maps using both average-pooling (avg) and max-pooling (max) operations simultaneously along the spatial axis. The spatial attention focused on finding where the information is located using global average-pooling and max-pooling along the channel axis followed by a convolution layer (kernel size = 7, number of filters = 1, stride = 1). The CBAM was implemented after each CNN layer on the branch with the smaller kernel size (the temporal branch).

  4. CNN-CBAM-LSTM: The CNN-LSTM model augmented with CBAM layers.

Three established open-source methods were used for comparison:

  1. Spectral power threshold-based detection: First, the 50 Hz power-grid noise was removed from the raw EEG with a notch filter, followed by a band-pass between 0.5 and 40 Hz with a Butterworth band-pass filter. A baseline threshold was calculated per night by averaging the power in the 0.75-4.5 Hz and 20-30 Hz bands from epochs that were manually scored as N1, N2, and N3 sleep stages.

  2. Standard deviation threshold-based detection (YASA Toolbox): "The YASA algorithm processes the whole recording at once, dividing it into small windows of predefined length. For each window, the standard deviation was first computed, and the resulting array was then log-transformed and z-scored. Windows with values exceeding the threshold were considered artifacts." Window length was set to 4 seconds.

  3. Heuristic-based 1D-CNN: Paissan et al. proposed a one-dimensional CNN architecture for detecting single-channel EEG artifacts consisting of a convolutional layer without downsampling, batch normalization, a rectified linear unit (ReLU) activation function, and global average pooling followed by two fully connected layers with 8 and 3 hidden units, respectively.

The CNN-CBAM model achieved the best classification performance: The CNN-CBAM model outperformed the other three deep learning architectures, achieving the highest classification metrics with an AUC of 0.88, a se of 0.81, and a sp of 0.86. This was achieved with a parameter count of 1,080,456 and inference time of 0.006 ms per epoch.

The ablation study results were:

  • CNN: AUC 0.73, se 0.66, sp 0.68

  • CNN-LSTM: AUC 0.77, se 0.66, sp 0.80

  • CNN-CBAM: AUC 0.88, se 0.81, sp 0.86

  • CNN-CBAM-LSTM: AUC 0.84, se 0.78, sp 0.82

Benchmark results:

  • Spectral power approach: se 0.35, sp 0.79 (unable to classify artifacts in REM sleep and Wake stages)

  • YASA: AUC 0.72, se 0.64, sp 0.69

  • 1D-CNN: AUC 0.85, se 0.76, sp 0.81

The authors note: "A major observation from our ablation study is that the CNN-CBAM model outperformed the CNN-CBAM-LSTM model, which incorporates recurrent layers... Adding a Bi-LSTM layer on top may introduce parameter redundancy and unnecessary architectural complexity, ultimately reducing performance, while also increasing inference cost - 0.057 ms versus 0.006 ms per 20-second EEG epoch for CNN-CBAM."

For localization, the attention mechanism of CNN-CBAM achieved a se of 0.61 and a sp of 0.63, with an optimal threshold at 0.66. The Jaccard index was 9.74% for CNN-CBAM and 5.52% for YASA, indicating CNN-CBAM achieving nearly twice the overlap accuracy of YASA.

The authors explain: YASA's performance was more significantly compromised by the initial misclassification of 20-second EEG epochs, which led to a higher number of false positives in the localization step. The confusion matrices revealed All positives: 6,202 for CNN-CBAM (True positives: 1,206; False positives: 4,996) versus All positives: 11,877 for YASA (True positives: 959; False positives: 10,918).

  1. All deep learning models outperformed traditional feature-based signal processing algorithms that rely on spectral power or standard deviation thresholds.

  2. The attention mechanism enabled granular identification of noise within the 20-second epochs and provided clear insight into which signal sections contributed to a positive artifact classification.

  3. The authors note: "The qualitative evaluation of the CNN-CBAM attention map showed that the high-amplitude attention regions closely matched the reference labels used in manual sleep scoring. However, this alignment was not perfect, yielding a localization se of 0.61."

  4. The authors acknowledge limitations: "our work incorporated only a single training run per model with a held-out test set. We therefore cannot rule out the possibility that the reported ranking within developed deep learning models is partly split-dependent given the limited number of participants and the class imbalance in artifact epochs."

  5. The method is suitable for deployment in quasi-real-time monitoring systems because it does not require prior information about individual subject baselines or manual sleep stage scoring.

  6. The authors note that "the ideal attention threshold used for reporting was obtained on the same segments predicted as artifact-contaminated, which may introduce a bias toward this specific dataset. Independent validation on an external dataset is required to confirm the generalizability of the selected threshold fully."

The authors conclude: "We proposed a deep learning framework for detecting artifacts in EEG data collected by wearable sleep devices. An ablation study revealed that the CNN-CBAM model outperformed the alternative CNN, CNN-LSTM, and CNN-CBAM-LSTM architectures, as well as the established open-source algorithms based on spectral signal processing. This research demonstrates a novel and effective approach for quasi-real-time artifact detection in raw, single-channel EEG signals. Crucially, the system eliminates the labour-intensive process of manual feature engineering. We further demonstrate that integrating attention maps with a single, tunable parameter provides a mechanism for artifact localization, enabling the rapid visualization of artifacts."

Improvements for AI systems

Based on the scientific paper, here are the specific improvements I can make to AI systems and what the improved system can do:

  • Improvement: Implement the two-branch convolutional architecture (small kernel for temporal detail, large kernel for frequency patterns) with the 1D-adapted Convolutional Block Attention Module (CBAM) applied after each convolutional layer on the temporal branch.

  • Specific change: Replace standard CNN layers with the described kernel sizes (12 and 100), apply channel attention (avg/max pooling → shared MLP with reduction ratio 8 → sigmoid) followed by spatial attention (avg/max pooling along channel axis → concatenation → 1D convolution with kernel 7 → sigmoid) sequentially.

  • Improvement: Add the activation mapping function F sum(A) = sum i=1 C A i p with p=4, applied after the last CBAM layer of the temporal branch.

  • Specific change: This produces a continuous temporal attention map that can be thresholded (optimal threshold 0.66) to localize artifacts within 20-second epochs at 4-second resolution.

  • Improvement: Use the ablation study results to avoid unnecessary complexity. The paper shows CNN-CBAM (AUC 0.88) outperforms CNN-CBAM-LSTM (AUC 0.84) despite fewer parameters (1.08M vs 1.43M) and 10x faster inference (0.006 ms vs 0.057 ms per epoch).

  • Specific change: Drop LSTM layers when CBAM is present; the attention mechanism already captures temporal dependencies effectively.

  • Improvement: Apply SMOTE oversampling to training/validation sets (not test set) to balance artifact/non-artifact classes (from 5% to 50% artifact epochs), then validate synthetic samples via PSD matching to ensure morphological plausibility.

  • Specific change: Use min-max scaling per epoch before SMOTE; verify synthetic samples preserve spectral characteristics (dominant frequency, spectral centroid, band powers).

  • Improvement: Select classification threshold by maximizing sqrt se times sp on validation set (not test set), with threshold search from 0 to 1 in 0.01 increments.

  • Specific change: This yields the reported se=0.81, sp=0.86 for CNN-CBAM; for localization, apply same optimization on predicted artifact segments to find threshold 0.66.

  1. Real-time artifact detection in single-channel mobile EEG at 128 Hz sampling, processing 20-second epochs in 0.006 ms (GPU) — suitable for IoMT edge devices or cloud deployment.

  2. Temporal artifact localization within epochs: identify which 4-second windows contain artifacts (se=0.61, sp=0.63, Jaccard=9.74%) using the attention map, enabling selective data rejection rather than discarding entire epochs.

  3. No manual feature engineering: operates directly on raw min-max scaled EEG, requiring only resampling to 128 Hz and 20-second segmentation.

  • Tunable sensitivity: Adjust the attention threshold (e.g., 0.66) to match clinical preferences; lower threshold increases recall, higher increases precision.

  • Visual interpretability: Generate attention maps alongside EEG traces (as in Figure 7) for semi-automated human review, reducing manual scoring workload.

  • Subject-independent: No per-night baseline calibration or manual sleep staging required (unlike spectral power approach), enabling plug-and-play deployment across users.

  • Sleep research: Automatically flag corrupted epochs for exclusion from sleep staging, generate data availability graphs, and provide pre-annotated artifact labels for human scorers.

  • Wearable EEG monitoring: Trigger real-time user alerts (e.g., adjust electrode band) when artifact density exceeds thresholds, improving data quality during recording.

  • Cloud-based analytics: Process large-scale multi-night recordings in parallel, providing quality metrics and artifact localization for downstream analysis.

  • Outperforms YASA toolbox (AUC 0.72 vs 0.88) and spectral power approach (se 0.35 vs 0.81).

  • Comparable to heuristic 1D-CNN (AUC 0.85) but adds localization capability that 1D-CNN lacks.

  • Maintains high specificity (0.86) while achieving sensitivity (0.81), critical for avoiding unnecessary data loss in sleep studies.

Abstract

Current methods for detecting artifacts in sleep EEG range from threshold-based algorithms to machine learning approaches, yet applications remain limited for single-channel mobile EEG. We propose a convolutional neural network (CNN) model incorporating a convolutional block attention module (CNN-CBAM) to detect and localize artifacts in sleep EEG using attention maps. We benchmarked this model against 6 other machine learning and signal processing approaches. We trained/tuned all models on 72 manually annotated EEG recordings obtained during home-based monitoring from 18 healthy participants with a mean (SD) age of 68.05 y (plus or minus 5.02). We tested them on 26 separate recordings from 6 healthy participants with a mean (SD) age of 68.33 y (plus or minus 4.08), which contained artifacts in 4% of epochs. CNN-CBAM achieved the highest area under the receiver operating characteristic curve (0.88), sensitivity (0.81), and specificity (0.86) among the tested approaches. Under the ideal choice of an attention threshold of 0.66, the attention maps from CNN-CBAM localized artifacts within detected artifact epochs with a sensitivity of 0.61 and specificity of 0.63. This work demonstrates the feasibility of automating artifact detection and localization in wearable sleep EEG.

Related papers