Lightweight In-Vehicle BCI via Blink-Event Gating: Single-Channel EEG Intent Recognition With Explainable Boosting

Abstract

Electroencephalography (EEG)-based brain–computer interfaces (BCIs) have emerged as a viable hands-free pathway for human–vehicle interaction in intelligent transportation systems. However, existing EEG-based paradigms face critical barriers to practical in-vehicle deployment. Endogenous approaches such as motor imagery require multi-channel hardware and prolonged calibration. Exogenous approaches such as P300 and SSVEP compete with the driving task for visual attention. Prevailing deep learning decoders lack the interpretability essential to safety-critical applications. Moreover, although blink-derived EEG signals have also supported assistive control, their integration as an intent-discrimination gate before in-vehicle actuation remains underexplored. To deal with these gaps, this study proposes a lightweight single-channel EEG-based BCI system that employs voluntary blink events as active control commands. An adaptive threshold algorithm based on median absolute deviation (MAD) first detects candidate blink events. Four temporally aggregated features, including blink intensity (<inline-formula> <tex-math notation="LaTeX">$E_{1s}$ </tex-math></inline-formula>), rolling standard deviation <inline-formula> <tex-math notation="LaTeX">$\sigma (E_{1s})$ </tex-math></inline-formula>, instantaneous change (<inline-formula> <tex-math notation="LaTeX">$\Delta E_{1s}$ </tex-math></inline-formula>), and active ratio (<inline-formula> <tex-math notation="LaTeX">$r_{active}$ </tex-math></inline-formula>), are then constructed. These features are classified by an Explainable Boosting Machine (EBM), whose one-dimensional shape functions provide transparent, auditable decision logic. The system is deployed on a Raspberry Pi 5 edge-computing platform, achieving microsecond-level inference latency within a 0.1 s decision cycle. Evaluation on a four-subject dataset of 4,728 samples yields 84.04% classification accuracy and an AUC of 0.902. A complementary leave-one-subject-out (LOSO) evaluation further supports the cross-subject applicability of the proposed classifier. The proposed EBM outperforms the fixed-threshold baseline (AUC = 0.671), the logistic regression baseline (AUC = 0.763), and the SVM baseline (AUC = 0.849). Real-vehicle control experiments with four unseen subjects further demonstrate 85% and 80% command accuracy for single-blink and double-blink commands, respectively. These results confirm the feasibility of an interpretable, low-cost, single-channel BCI for supplementary in-vehicle interaction. This provides a basis for trustworthy driver-intent decoding in intelligent cockpit applications.