Integrating contextual cues from the driver, cabin, and surrounding environment is crucial for enhancing semi-automated vehicle safety. Current systems often rely on video streams to capture driver behaviours and environmental conditions, yet face challenges in low-light settings and privacy intrusion. In this study, we investigated multimodal physiological signals and Kernelized Convolutional Neural Networks (K-CNN) for assessing driver behaviour. For this, electrocardiography (ECG) and respiration (RSP) data were collected from 15 healthy volunteers wearing smart shirts in two driving scenarios. Signals were sampled at 256Hz and 128Hz, respectively. From this data we have extracted time-domain, frequency-domain and non-linear features using Neurokit2 toolkit. These features are fed to the K-CNN framework where task specific representation is extracted using Radial Basis Function, Linear, Polynomial Degree 2 and Polynomial Degree 3 kernels and these representation fed to 1D CNN to extract temporal features and these features are fused to detect driver behaviour using sigmoid. The proposed approach is able to classify driver behaviour states using multimodal data, achieving an average accuracy of 70.54% and an average F-score of 62.22% for linear kernel. Textile-based physiological data proved more reliable in classifying car driver behaviour states using K-CNN than with conventional machine learning techniques.

Prolonged driver inattention significantly increases the risk of traffic accidents, necessitating continuous physiological monitoring for improved road safety and driver wellbeing. This study proposes a Self-Attention-based Hybrid Quantum Neural Network (SAHQNN), a novel framework integrating trainable Parameterized Quantum Circuits (PQCs) with selective self-attention for automated detection of phone call distraction-induced cognitive and emotional driver inattention from multimodal wearable biosignals. Multimodal physiological signals consisting of single-lead Electrocardiogram (ECG, 256 Hz) and Respiration (RSP, 128 Hz) were acquired from $N=20$ participants under Normal and Distracted-Inattention driving conditions using a textile wearable smart shirt. The Distracted-Inattention condition was induced through a hands-free phone call that simultaneously imposed cognitive load, emotional arousal, and secondary task engagement on the driver through active questioning, reflecting the multidimensional nature of driver inattention beyond speech activity alone. Multi-domain features reduced via Random Forest Feature Importance (RFF) were angle-encoded into 10-qubit PQCs with 20 trainable variational $R_Y\left(\theta \right)$ gates optimized via the parameter-shift rule, establishing inter-qubit correlations through Hadamard and Controlled-NOT (CNOT) entanglement layers. Pauli-Z measurement outputs were processed through a selective self-attention layer, producing Attentive Quantum Features (AQF) that were subsequently classified by fully connected layers. The proposed SAHQNN achieves 77.50% accuracy and 70% weighted F-measure under Leave One Subject Out Cross-Validation (LOSOCV), with statistically significant improvements over all standard classical baselines ($p<0.001$, Cohen’s $d>1.5$) and 61% fewer parameters than the strongest classical competitor, demonstrating the feasibility of parameter-efficient trainable hybrid quantum neural networks for subject-independent driver inattention detection.

Low-cost webcam gaze estimation is increasingly proposed as a front-end sensor for self-monitoring of attention and visual well-being, yet average accuracy hides where and for whom such a sensor is reliable. Using only three public gaze datasets (MPIIFaceGaze, Columbia Gaze, and RT-GENE; no new data collection), we make three contributions. First, taking a naive sample-wise split (shuffling samples, not subjects) that places every one of 84 subjects across train/validation/test as a deliberate reference baseline, we show apparent error rising from 4.47 deg on seen subjects to 11.63 deg on subject-disjoint unseen subjects (each a five-seed mean; 95% CI: +/−0.11 deg and +/−4.64 deg, respectively), with an optimism of +7.15 deg (~62%); every naive seed lies far below every subject-disjoint seed, and because the unseen estimate varies widely across seeds, its magnitude should be read as indicative rather than precise. Second, we audit reliability per subject and head pose stratum without demographic labels, revealing a 4.9× spread across users and a Monte Carlo dropout uncertainty that provides a partial, borderline per-subject reliability signal (Pearson r = 0.48; Spearman rho = 0.50, with a bootstrap interval including zero); bias-only recalibration lowers error only modestly (subject-macro: 11.79 to 11.50 deg) and does not help the least-reliable users. Third, we convert this measured reliability profile into a per-user decision (trust/recalibrate/abstain) using objective-weight multi-criteria decision making (averaged CRITIC and entropy weights with TOPSIS; the largest weight falls on MC-dropout uncertainty at 0.42, and then pose sensitivity at 0.23); the decision layer disagrees with a single-metric threshold for 35% of users (six of 17) while remaining stable under +/−20% weight perturbation (no action changes). This study provides an exploratory, retrospective, public-data-only implementation of a per-user reliability decision scaffold for low-cost gaze-based self-monitoring.