Cross-modal cognitive workload estimation in driving simulation: CCI–pupil coherence, vestibular mediation, and a multimodal workload index across five washout algorithm conditions

This paper investigates the cross-modal relationship between the biceps–triceps co-contraction index (CCI)—derived from surface EMG of the upper limb—and pupil dilation SD (PD-SD)—derived from binocular pupillometry—across five washout algorithm conditions in a high-fidelity driving simulator. Unlike prior multimodal workload studies that combine simultaneous physiological channels under fixed task conditions, this work exploits five distinct motion cueing conditions to test whether spectral coherence between CCI and pupil dilation is modulated by vestibular input quality—a dimension not previously examined in driving simulator research. The primary aim was to determine whether spectral coherence between these two physiological channels, particularly in the 0.1–0.3 Hz locus coeruleus–norepinephrine (LC-NE) band, varies monotonically with washout fidelity and whether vestibular sway power mediates this relationship. Cognitive workload during driving is shaped not only by the demands of the primary task but by the quality of sensory feedback provided by the simulator’s motion cueing system. Pupil dilation tracks attentional resource engagement via the LC-NE system, while upper-limb EMG co-contraction reflects the motor system’s response to perturbation predictability. The ascending vestibular–LC projection and the vestibulo-spinal pathway provide anatomical substrates linking platform sway to both signals. We hypothesise that these shared pathways create a mechanistic three-way linkage between platform sway, neuromuscular tone, and pupil dilation whose expression in spectral coherence has not previously been examined in the context of washout algorithm selection. Data were drawn from the SAAM driving simulator experiment (ENSAM ParisTech; N = 28 participants, 23M/5F, mean age 29.0 ± 4.6 years) across five conditions: Static, Classic (fixed high-pass), A1 (MRAC adaptive), H1 (Hybrid MRAC + LQR), and O1 (optimal LQR/MPC). Time-resolved CCI was derived from 13-channel surface EMG using a 3-second sliding Hanning window. Pupil diameter was recorded at 60 Hz via SMI iView X; it was first downsampled from 60 Hz to 10 Hz (3-sample block averaging) and smoothed with a 5-point Gaussian kernel (σ = 1 sample at 10 Hz), then resampled to a 1 Hz common grid for coherence estimation. Magnitude-squared coherence C 2 ( f ) between CCI and pupil dilation was computed in the 0.1–0.3 Hz band using Welch’s method with Bonferroni correction. Vestibular mediation was assessed via Baron–Kenny partial correlation with 5000-iteration bootstrapping. A Multimodal Workload Index (MWI) was constructed from CCI, PD-SD, and vestibular sway power with SHAP-derived weights. CCI–pupil coherence followed the fidelity hierarchy exactly: C 2 (Classic) = 0.67, C 2 (A1) = 0.48, C 2 (H1) = 0.39, C 2 (O1) = 0.22—a 67% attenuation from fixed-gain to receding-horizon optimal control. Vestibular-band sway PSD mediated 62% of the CCI–pupil coupling (bootstrapped 95% CI: [0.47, 0.75], p = 0.008). The Classic condition exhibited a workload dissociation: highest CCI (0.85) but lowest mean pupil diameter (3.25 mm), undetectable by single-modality analysis. H1 produced a uniquely stable attentional profile (high mean pupil = 4.27 mm, lowest SD = 0.61 mm, highest kurtosis = 6.62). The MWI achieved test–retest ICC(2,1) = 0.84 (95% CI: 0.71–0.92) and significantly discriminated the four active conditions (Kruskal–Wallis H = 14.7, p = 0.002). CCI–pupil spectral coherence in the 0.1–0.3 Hz band is a sensitive cross-modal index of cognitive workload in driving simulation, with its magnitude consistent with modulation by washout algorithm fidelity via a vestibular–LC pathway (exploratory, condition-level analysis). The Classic workload dissociation underscores the inadequacy of single-modality physiological assessment. The MWI is proposed as a standardisable physiological fidelity criterion for ISO-referenced driving simulator evaluation, subject to external validation with larger, multi-laboratory samples. Practitioners evaluating or selecting driving simulators for cognitive workload research, driver monitoring system development, or safety-critical training should employ the MWI—combining upper-limb EMG co-contraction, pupillometric variability, and vestibular sway power—as a three-feature physiological benchmark. Classic fixed-gain washout produces a workload dissociation that will mislead single-modality metrics: EMG alone overestimates workload while pupillometry alone underestimates it. Simulators with at minimum an MRAC-class adaptive washout algorithm are required for valid physiological workload measurement. The MWI framework is directly applicable to ISO-referenced simulator certification and to the calibration of real-time gaze-contingent and EMG-contingent ADAS systems.