Individualized stress detection using an unmodified car steering wheel

Abstract

In-car passive stress sensing could enable the monitoring of stress biomarkers while driving and reach millions of commuters daily (i.e., 123 million daily commuters in the US alone). Here, we present a nonintrusive method to detect stress solely from steering angle data of a regular car. The method uses inverse filtering to convert angular movement data into a biomechanical Mass Spring Damper model of the arm and extracts its damped natural frequency as an approximation of muscle stiffness, which in turn reflects stress. We ran a within-subject study (N = 22), in which commuters drove a vehicle around a closed circuit in both stress and calm conditions. As hypothesized, cohort analysis revealed a significantly higher damped natural frequency for the stress condition (P = .023, d = 0.723). Subsequent automation of the method achieved rapid (i.e., within 8 turns) stress detection in the individual with a detection accuracy of 77%.

Figure 1

Individualized driver stress detection method using an unmodified car steering wheel.

Figure 2

Results of the stressor validation analysis to establish a ground-truth assessment of whether participants were more stressed during stress exposure compared to calm conditions. (a) Boxplots of normalized and baseline corrected subjective and psychophysiological metrics. Permutation probability P values are corrected for multiple-comparisons with ***P < 0.001; **P < 0.01, and *P < 0.05. (b) Permutation results with adjusted probability p values after Bonferroni correction, and Cohen’s d as effect size measure reference.

Figure 3

Results for cohort analyses. (a) LPC model fit with minimum, median, maximum variance of the residual errors across participants. Overall LPC models showing good quality of fit in all three examples. (b) Shows median values and boxplots of damping frequencies, with “c” = calm drive, “s” = stress drive, and ‘d” = Cohen’s d. P values need to fall below 0.0125 to be significant due to Bonferroni correction for multi-comparison.

Figure 4

Variance of residual error and average number of turn segments across the iterated upper bound segmentation parameters, with decreasing variance of residual error (i) and increasing average number of turn segments (ii) across upper bound values. The peak in average number of turn segments is at 120° upper bound (iii). Please find the definition of upper bound in the Methods section.

Figure 5

Steering wheel stress sensing method. (a) A typical steering wheel movement and corresponding LPC trajectory. (b) Steering angle data for one participant with negative values for counter clockwise rotation; and, (c) processed steering angle data along with the selected turn segments.

Figure 6

Experimental methodology. Procedure of the experiment (a), in which frequent commuters drove in a regular car (b), within a large parking garage (c), on a fixed 0.65 miles driving course that included eight turns per lap (d).