Predicting vehicle operators’ neck discomfort under vibration using head posture data: A dynamic two-stage attention-based method

Abstract

Vehicle vibration degrades interaction with In-Vehicle Information Systems (IVIS) by increasing neck discomfort and head-posture instability. We ran three laboratory experiments manipulating interface context while applying 0–2.5 Hz (Hz) vertical vibration. Participants rated neck discomfort relative to vibration on the Body Parts Perceived Level of Discomfort (BPPLD) scale; head-angle root mean square (RMS) was derived from an inertial measurement unit (IMU). Linear mixed-effects models showed monotonic vibration-driven increases in discomfort and head-angle RMS. Building on these findings, we implemented a deep learning method—a dynamic two-stage attention network that fuses vibration and tri-axial head-posture features through vibration-gated multi-head self-attention—and applied it to binary high-discomfort prediction for ergonomic state monitoring of vehicle operators. The model was benchmarked under identical, strictly leakage-free protocols against mainstream deep learning baselines (long short-term memory and gated recurrent unit networks, one-dimensional convolutional neural networks, graph neural networks, tabular residual networks). Under block-wise forward prediction (training on 0–2.0 Hz, testing on the unseen 2.5 Hz extreme condition), the proposed network achieved 71.79% accuracy, 0.7052 balanced accuracy, 0.7010 macro-averaged F1-score, and 0.8540 area under the receiver operating characteristic curve (AUC)—outperforming all compared baselines on all four metrics, and yielding the best discrimination and the highest high-discomfort sensitivity (recall 0.896) of all non-degenerate models. Under leave-one-subject-out (LOSO) cross-validation with fold-isolated normalization and training-side threshold selection, it achieved 69.20% accuracy and 0.6692 AUC, evidencing cross-subject generalization without data leakage. Calibrated risk scores with dynamically optimized decision thresholds support sensitivity-prioritized warning policies for adaptive IVIS on unpaved roads.