Multimodal driving style classification in a six-DOF driving simulator: Decomposing washout algorithm effects from driver-specific patterns in EMG, pupillometry, and vehicle dynamics

Driving style classification in simulator-based studies is substantially confounded by the motion cueing washout algorithm, which independently alters the driver’s neuromuscular and cognitive responses. This paper quantifies this confound and proposes a condition normalisation procedure using driving simulator data. The dataset includes N = 28 participants (23 males, five females; mean age 29.0 ± 4.6 years; licence experience 10.0 ± 3.4 years) completing seven washout conditions—Classic (C), MRAC-A1/A2, Hybrid-H1/H2, LQR-O1/O2—in a within-subject counterbalanced design, with simultaneous recording of (i) 13-channel upper-limb surface EMG at 11.2 Hz, (ii) binocular pupillometry at 60 Hz, (iii) six-DOF motion platform signals at 20 Hz, and (iv) vehicle IMU dynamics at 20 Hz. Four modality-specific feature vectors were constructed: 106 EMG features (RMS, spectral onset index, CCI, bilateral ratio), seven pupillometric features (mean diameter, SD, blink rate, band powers), 12 platform features (acceleration RMS per axis, sway–yaw Lissajous coupling index), and 14 vehicle dynamics features (lateral acceleration statistics, yaw rate RMS, spectral band powers). Within-subject z -score normalisation removes the washout main effect from the physiological feature space. Support Vector Machine, Random Forest, and XGBoost classifiers were evaluated under leave-one-out cross-validation on raw features and normalised residuals. XGBoost achieved F 1 = 0.891 on condition identification and F 1 = 0.712 on driving style from normalised features, compared to F 1 = 0.431 without normalisation. SHAP attribution identified the right-arm co-contraction index (CCI_R), pupil dilation standard deviation, and lateral sway RMS as the top discriminating features. The 1–2 Hz EMG spectral onset peak is proposed as a spectral fidelity index for retroactive diagnosis of algorithm contamination in archival datasets. These results establish multimodal, condition-aware feature engineering as a prerequisite for valid driving style classification in driving simulators.