PaperlayerView exact revision on arXiv

Source-linked AI summary

A Compact and Interpretable Convolutional Neural Network for Cross-Subject Driver Drowsiness Detection from Single-Channel EEG

Jian Cui, Zirui Lan, Yisi Liu, Ruilin Li, Fan Li, Olga Sourina, Wolfgang Mueller-Wittig

arXiv:2106.00613v1eess.SPcs.HCcs.LGcs.NE

TL;DR

Cross-subject discovery of shared EEG features remains underexplored, despite EEG’s association with drowsiness. The paper uses a compact CNN with GAP to visualize learned signal regions, reports that it outperforms conventional and deep-learning baselines, and identifies interpretable patterns associated with drowsy and alert states.

  • Problem

    Discovering and visualizing shared EEG features across subjects remains underexplored, although EEG oscillations are associated with drowsiness.

    Source: §II · pp. 2–3

  • Method

    A compact CNN with a GAP layer uses CAM to visualize the single-channel EEG regions contributing to classification.

    Source: §VII · p. 15

  • Results

    The model outperformed conventional and deep-learning baselines and identified Alpha spindles and Theta bursts as drowsiness evidence, with muscle artifacts and sensor drifts used to recognize alertness.

    Source: §VII · p. 15, §VIII · p. 16

  • Takeaways & Limitations

    The results illustrate a potential use of CNNs for discovering neurophysiological patterns in EEG signals.

    Source: §VIII · p. 16

    Behavioral performance measures may not faithfully reflect mental state in some circumstances, complicating the fidelity of ground-truth labels.

    Source: §VII · p. 16

Abstract

from arXiv · show

Driver drowsiness is one of main factors leading to road fatalities and hazards in the transportation industry. Electroencephalography (EEG) has been considered as one of the best physiological signals to detect drivers drowsy states, since it directly measures neurophysiological activities in the brain. However, designing a calibration-free system for driver drowsiness detection with EEG is still a challenging task, as EEG suffers from serious mental and physical drifts across different subjects. In this paper, we propose a compact and interpretable Convolutional Neural Network (CNN) to discover shared EEG features across different subjects for driver drowsiness detection. We incorporate the Global Average Pooling (GAP) layer in the model structure, allowing the Class Activation Map (CAM) method to be used for localizing regions of the input signal that contribute most for classification. Results show that the proposed model can achieve an average accuracy of 73.22% on 11 subjects for 2-class cross-subject EEG signal classification, which is higher than conventional machine learning methods and other state-of-art deep learning methods. It is revealed by the visualization technique that the model has learned biologically explainable features, e.g., Alpha spindles and Theta burst, as evidence for the drowsy state. It is also interesting to see that the model uses artifacts that usually dominate the wakeful EEG, e.g., muscle artifacts and sensor drifts, to recognize the alert state. The proposed model illustrates a potential direction to use CNN models as a powerful tool to discover shared features related to different mental states across different subjects from EEG signals.

II. RELATED WORKS

Prior work links EEG rhythms with drowsiness and explores conventional and deep-learning recognition, but cross-subject feature discovery and visualization remain underexplored. The study uses a simulated-driving EEG dataset collected from 27 participants.

II. RELATED WORKS

  • EEG oscillations have been associated with drowsiness, including higher low-frequency power during night driving and increased lower-alpha power when sleepy subjects struggle to stay awake.The cited examples include Torsvall, Lal et al., and Klimesch et al..Source: §II · p. 2
  • Conventional single-channel methods engineer spectral, entropy, or wavelet features; Ogino et al. reported 72.7% accuracy using PSD features with SWLDA and SVM.Other approaches use deviation from alert-state spectra, wavelet features, or individual alpha frequency.Source: §II · p. 2
  • Deep-learning work spans single-channel EEG sleep, abnormality, attention, and drowsiness classification using CNNs and recurrent architectures.Examples include CNN models and a recurrent model reporting Cohen’s Kappa of ~0.73 on 19 subjects.Source: §II · pp. 2–3
  • Cross-subject discovery and visualization of shared EEG features remain underexplored, motivating a compact CNN designed to reveal learned patterns.The authors contrast this aim with prior deep-learning work that treats models mainly as black-box classifiers.Source: §II · p. 3

A. Dataset description

  • The dataset contains 62 EEG recordings from 27 participants, who completed a monotonous simulated-driving task while responding to lane departures.Signals were sampled at 500 Hz using a 30-electrode EEG cap and two reference electrodes.Source: A. Dataset description · p. 3

B. Data extraction

The study preprocesses and filters EEG data, labels samples using both local and longer-term reaction times, then balances subjects and classes. The resulting set contains 2,022 samples from 11 subjects.

B. Data extraction

  • The EEG was down-sampled to 128 Hz, and 3-second Oz-channel samples were extracted before each lane-deviation onset.Each sample contains 384 points from one channel; eye-blink contamination was manually removed and ocular and muscular artifacts were removed with EEGLAB’s AAR plug-in.Source: B. Data extraction · p. 3
  • Samples were labeled using local reaction time and a 90-second average global reaction time relative to each session’s alert baseline.Alert required both measures below 1.5 times alert-RT; drowsiness required both above 2.5 times alert-RT, excluding the moderate transition range.Source: B. Data extraction · p. 3
  • The researchers discarded sessions with fewer than 50 samples in either class, chose the most balanced session per subject, and balanced classes within sessions.For within-session balancing, they retained majority-class samples with the shortest alert or longest fatigue local reaction times.Source: B. Data extraction · p. 4
  • Balancing was intended to limit subject-favoring predictions and encourage discrimination between alert and drowsy samples.The three filtering steps exclude highly skewed sessions and loosely balance subjects as well as classes.Source: B. Data extraction · p. 4
  • The study obtained 2,022 samples in total from 11 different subjects; Figure 1 displays extracted alert and drowsy samples across experimental time, with response time on the y-axis.Source: Figure 1 · p. 4, B. Data extraction · p. 5

IV. METHOD

The method uses a compact one-dimensional CNN for single-channel EEG classification. Its Global Average Pooling layer supports CAM-based interpretation while reducing parameters.

IV. METHOD

  • A 64-point convolutional filter is applied with 32 channels, followed by batch normalization and ELU activation.The filter length is half the 128-Hz sampling rate; the authors report that batch normalization accelerated convergence and improved network stability in preliminary tests.Source: A. Network design · p. 5
  • The model uses Global Average Pooling (GAP) instead of fully connected layers to link feature maps to classes and enable CAM visualization.GAP also reduces model parameters, acts as a structural regularizer, and is described as robust to spatial translations that may help detect EEG microstructures.Source: A. Network design · p. 5, Figure 2 · p. 6
  • The compact design is intended for relatively small datasets, with about 2,000 samples, and permits direct interpretation of classifications using CAM.The authors also report that GAP accelerates convergence.Source: A. Network design · p. 5

B. Visualization with the CAM method

The CAM procedure combines class-specific feature-map activations to identify signal regions contributing to a decision. The resulting map is aligned to the input and converted into a normalized heatmap.

B. Visualization with the CAM method

  • The CAM method visualizes discriminative input regions as evidence for classification, allowing learned patterns to be examined across subjects.In Equation (2), the dense-layer output is the network’s final activation for class c.Source: B. Visualization with the CAM method · p. 6
  • The class activation map combines feature-map activations weighted for the target class, and its entries sum to the class’s final activation.The map has length 321 for the described network and represents regions contributing to classification.Source: B. Visualization with the CAM method · p. 6
  • For the single-convolution model, the activation map is shifted toward the input’s center to align it more accurately with contributing signal regions.The procedure uses a 64-point kernel and a 384-point input; the final heatmap is normalized, and negative class-map activations are muted with ReLU.Source: B. Visualization with the CAM method · pp. 6–7

A. Comparison methods

The paper compares its model with two deep-learning classifiers and conventional classifiers using band-power features. The deep-learning baselines are adapted to single-channel EEG where needed.

A. Comparison methods

  • The comparison includes state-of-the-art deep-learning and conventional methods, with Deep CNN adjusted for the study’s sampling rate.The first convolution’s kernel and stride change from (1, 4) and (1, 2) to (1, 6) and (1, 3), respectively.Source: A. Comparison methods · p. 7
  • Deep CNN classifies attention versus non-attention from single-channel EEG frequency components.Its architecture uses convolutional and max-pooling layers followed by dropout, a fully connected layer, and Softmax.Source: A. Comparison methods · p. 7
  • EEGNet combines temporal and depthwise convolutions with separable convolution before classification.The implementation removes spatial layers designed for multichannel EEG and uses an EEGNet-8,2-based configuration.Source: A. Comparison methods · p. 7
  • The conventional baselines classify four relative band-power features extracted from Oz using Welch’s method.The tested classifiers are DT, RF, KNeighbors, GaussianNB, LR, LDA, QDA, and SVM.Source: A. Comparison methods · p. 7

B. Implementation details

The experiments use Python and PyTorch for the proposed model, with TensorFlow/Keras for deep-learning baselines and sklearn for conventional baselines.

B. Implementation details

  • The proposed model uses PyTorch, while Deep CNN and EEGNet use Keras with TensorFlow 1.13.1 and conventional baselines use sklearn.Experiments run on Windows 10 with an NVIDIA GeForce GTX 1080; Adam uses a learning rate of 0.01 and other stated default parameters.Source: B. Implementation details · p. 8

VI. RESULTS

In leave-one-subject-out evaluation, the proposed model outperforms the deep-learning and conventional baselines, reaching 73.22% mean accuracy. Ablations indicate that ELU and batch normalization matter, while GAP supports faster convergence.

VI. RESULTS

  • Evaluation holds out each of 11 subjects in turn and averages results across 10 random initializations per subject and epoch.This yields 110 folds for each epoch.Source: A. Cross-subject classification accuracy · p. 8
  • The proposed model reaches 73.22% after six epochs and remains above 71% through 50 epochs, outperforming Deep CNN and EEGNet overall.EEGNet peaks at 70.48% after three epochs before settling around 67.5%; Deep CNN takes longer to converge to a lower accuracy.Source: A. Cross-subject classification accuracy · p. 8, Figure 3 · p. 8
  • 73.22% is the proposed model’s highest mean accuracy among nine methods, ahead of SVM at 69.27% and LR at 67.77%.Its mean accuracy is significantly higher than DT, RF, KNeighbors, GaussianNB, and QDA, but not significantly different from LR, LDA, or SVM.Source: A. Cross-subject classification accuracy · p. 8, Table 2 · p. 9
  • From Figure 4, removing the ELU activation layer from the original model drops average accuracy by more than 20%, while batch normalization accounts for around 6% of average classification accuracy.Source: A. Cross-subject classification accuracy · p. 9
  • GAP reaches its 73.22% peak after six epochs, whereas AvgPool80 peaks at 73.85% after 40 epochs; their highest mean accuracies are not significantly different.AvgPool40 takes even longer to converge; the GAP layer is associated with faster convergence.Source: A. Cross-subject classification accuracy · pp. 9–10, Figure 4 · p. 9

B. Visualization on the learned patterns

The model identifies two recurring drowsy-sample patterns: alpha spindles and Theta–Delta bursts. For alert samples, it uses high-frequency muscle activity and large-amplitude sensor drifts; visualizations of Subject 2’s errors suggest that mixed or atypical features can complicate classification.

B. Visualization on the learned patterns

  • High-likelihood drowsy samples show alpha spindles or Theta–Delta bursts, which the model uses as evidence of drowsiness.The alpha spindles can extend across cortical regions, while the bursts may occur in separate signal segments.Source: B. Visualization on the learned patterns · p. 10
  • Figure 5 shows correctly classified, high-likelihood drowsy examples, with Oz signals, CAM-highlighted regions, relative band power, and all-channel EEG panoramas.The subject, sample label, and alert and drowsy likelihoods are shown above each example.Source: Figure 5 · p. 11
  • High-likelihood alert samples commonly contain Beta-band activity from muscle tension or large-amplitude, multichannel sensor drifts.The reported drifts are distinguished from typical Delta waves because they occur simultaneously across channels.Source: B. Visualization on the learned patterns · p. 12
  • In Subject 2’s first misclassified drowsy sample, Beta activity supports an alert prediction while obvious alpha spindles or Theta–Delta bursts are absent.The paper notes that this sample lacks the drowsiness-related patterns observed in other examples.Source: B. Visualization on the learned patterns · p. 13
  • A misclassified alert sample with both Alpha and Beta activity has a higher drowsy-label likelihood than a sample with Beta activity alone, but is still classified as drowsy.The Alpha spindles in the mixed-feature sample are visible in frontal-central regions and weaken toward occipital channels.Source: B. Visualization on the learned patterns · pp. 13–14
  • The authors infer that Subject 2’s low accuracy is partly associated with drowsy-labeled samples that have high Beta activity and lack obvious drowsiness-related features.They also suspect that some alert labels may be unreliable when visualized signals contain strong drowsiness-related features.Source: B. Visualization on the learned patterns · p. 14

VII. DISCUSSION AND FUTURE WORKS

The discussion uses CAM visualizations to interpret the CNN’s learned EEG features and considers them for examining errors and labels. It also notes that the current evaluation setup and dataset constrain deployment and generalization.

VII. DISCUSSION AND FUTURE WORKS

  • The GAP layer enables CAM visualization of local input regions, and the authors report that it also helps testing accuracy converge faster.They note Grad-CAM as an alternative for generating activation maps with a broader range of CNN models.Source: §VII · p. 15
  • The model associates alpha spindles and Theta bursts with drowsiness, and muscle artifacts and sensor drifts with alertness.The authors describe these as biologically explainable features and wakeful-EEG artifacts, respectively.Source: §VII · p. 15
  • Visualizations of some wrongly classified samples lead the authors to suspect that behavior-based ground-truth labels may not faithfully reflect mental states in certain circumstances.They identify incorporating network explanations into labeling as a possible alternative to relying only on hard-coded behavioral or performance thresholds.Source: §VII · p. 16
  • The model is limited to offline analysis because evaluation batches all test-subject samples for batch normalization, and it has been tested only on Oz from one dataset with 11 subjects.The authors report around a 6% accuracy drop when batch normalization is removed and propose exploring other calibration approaches and broader testing.Source: §VII · p. 16

VIII. CONCLUSION

The conclusion presents a compact CNN for cross-subject drowsiness recognition from single-channel EEG. Its GAP layer supports visualization of classification-relevant signal regions, and the reported tests outperform conventional and deep-learning baselines while identifying interpretable features.

VIII. CONCLUSION

  • The cross-subject public-dataset test found that the compact CNN outperformed conventional and deep-learning baselines and localized interpretable micro-structures such as Alpha spindles.The authors conclude that CNNs may help discover neurophysiological patterns in EEG signals.Source: §VIII · p. 16

APPENDIX

The appendix table reports how many samples can be extracted from each raw-data session under the stated criteria and identifies the sessions used in the study.

APPENDIX

  • Table A1 lists extractable sample counts by session and identifies the 11 session indices used in the study.The caption refers to the extraction criteria described in Section III.B.Source: Table A1 · p. 19
Loading 2106.00613v1…