AI Enhanced Signal Processing Pipeline For Consumer Grade EEG - Based Brain Computer Interface
Driver fatigue and cognitive vigilance degradation constitute leading causes of severe vehicular collisions worldwide. While electroencephalography (EEG) offers high-temporalresolution neurophysiological correlates of cognitive fatigue, practical real-world adoption is severely hindered by three fundamental challenges: the physical invasiveness of high-density wholescalp electrode montages, low signal-to-noise ratios (SNR) caused by motion artifacts and dry-electrode impedance drift, and high inter-subject neurodynamic variability. This paper introduces Our approach, a lightweight, fully integrated deep learning framework designed for real-time driver vigilance decoding using exclusively five frontal EEG channels (FP1, FPZ, FP2, FZ, FC1) strategically aligned with non-hair-bearing wearable headbands. The pipeline integrates a 1D U-Net based Signal Enhancement Network (SEN) that reconstructs uncorrupted temporal waveforms in the raw signal domain, an EEGNet-inspired depthwise separable convolutional backbone for spatial-temporal filtering, a two-layer Bidirectional Long Short-Term Memory (BiLSTM) network for micro-state transition modeling, and a temporal self-attention mechanism that dynamically scores the diagnostic relevance of sub-epoch intervals. Evaluated across 162,817 two-second epochs from the 23-subject SEED-VIG driving benchmark under strict Leave-One-Subject-Out CrossValidation (LOSO-CV), Our approach attains a cross-subject classification accuracy of 79.70% ± 8.02%, an Area Under the ROC Curve (AUC-ROC) of 0.8168, and a macro F1-score of 0.7693, requiring only 258,707 trainable parameters. Extensive noise-injection stress testing confirms that the integrated SEN maintains an accuracy of 81.13% even under 50% additive Gaussian noise (6 dB SNR), while executing in under 4.2 ms on consumer edge hardware. The proposed architecture demonstrates that robust, subject-invariant vigilance monitoring is achievable from few-channel frontal EEG without compromising real-time performance.
Authors
- Deepthi Rapeti
- Altamash Shaikh
- Aniska Bachar
Institutions
- Thakur College of Engineering and Technology (IN)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-17
- DOI
- https://doi.org/10.5281/zenodo.24302472
- Primary Topic
- EEG and Brain-Computer Interfaces
- Type
- preprint