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

TL;DR

This paper tackles artifact detection in single-channel wearable sleep EEG by introducing a CNN-based model with a convolutional block attention module (CNN-CBAM) to both classify artifacts and localize their temporal occurrence. Four deep learning architectures (two CNN variants and their CBAM-augmented counterparts) are benchmarked against three open-source baselines on a home-recorded dataset of 98 nights from 24 older adults, with artifact labels at 4-s granularity. CNN-CBAM achieves the best artifact detection performance (AUC $=0.88$, se $=0.81$, sp $=0.86$) and provides artifact localization via attention maps, achieving se $=0.71$ and sp $=0.67$ at an optimized threshold. The results demonstrate feasibility for real-time artifact detection and localization in wearable sleep EEG, with implications for automated data quality control and edge/cloud deployment, though broader validation and multi-class artifact modeling are needed.

Abstract

Artifacts in the electroencephalogram (EEG) degrade signal quality and impact the analysis of brain activity. Current methods for detecting artifacts in sleep EEG rely on simple threshold-based algorithms that require manual intervention, which is time-consuming and impractical due to the vast volume of data that novel mobile recording systems generate. We propose a convolutional neural network (CNN) model incorporating a convolutional block attention module (CNN-CBAM) to detect and identify the location of artifacts in the sleep EEG with attention maps. We benchmarked this model against six 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 ($\pm$5.02). We tested them on 26 separate recordings from 6 healthy participants with a mean (SD) age of 68.33 y ($\pm$4.08), with 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) when compared to the other approaches. The attention maps from CNN-CBAM localized artifacts within the epoch with a sensitivity of 0.71 and specificity of 0.67. This work demonstrates the feasibility of automating the detection and localization of artifacts in wearable sleep EEG.

Figures (7)

• Figure 1: Artifact detection and localization from raw sleep EEG using a convolutional neural network model with an attention mechanism.
• Figure 2: Deep learning models used in a benchmarking study: a) baseline CNN and CNN-CBAM models, and b) CNN-LSTM and CNN-CBAM-LSTM models. The convolution block attention module (CBAM) only follows each convolutional layer on the temporal branch in CNN-CBAM and CNN-CBAM-LSTM models. Dropout was only performed during training with a rate of 0.5 (blue). The attention mapping was performed after the model's last CBAM layer (red), which was selected based on the benchmarking analysis of four deep learning models.
• Figure 3: The convolutional block attention module (CBAM) layer highlights the sequential connection of the channel- and spatial-wise attention in a convolutional neural network. (1) Features from the previous layer are squeezed along the spatial axis to obtain the inter-channel relationship, resulting in a channel-wise output. The multiplication of the input features from the previous layer and the resulting channel-wise output serves as the input for (2) where a spatial attention map is generated by squeezing the input along the channel axis.
• Figure 4: Implementation schema of (a) channel-wise attention and (b) spatial-wise attention in the convolutional block attention module (CBAM). (a) The global average- and max-pooling are applied to the input features to capture channel-wise dependencies. Two fully connected (FC) layers form a bottleneck, and rectified linear unit (ReLU) functions are used in between to create the non-linearity. The channel-wise output is created by adding outputs from both FC layers following a sigmoid activation. (b) The average- and max-pooling are applied and concatenated to capture spatial dependencies. A convolution layer (Conv) is then applied, followed by a sigmoid function to generate the spatial-wise output.
• Figure 5: Aggregated receiver operating characteristic (ROC) and the reported area under the curve (AUC) for our four deep learning models and three open-source algorithms. The dots on each ROC curve denote the best operating point with a balanced trade-off between $se$ and $sp$ as reported in Table \ref{['table:2']}. The dashed line indicates the performance of a uniform random guess classifier.