*3.2. Skin Conductivity*

The variation of skin conductivity is linked to the sympathetic nervous system [62]. When the driver has high stress, the activity of sweat glands is triggered by postganglionic sudomotor fibers. The result is a change in the skin conductivity response (SCR) that can be measured by applying a low constant voltage. The SCR amplitude can be used as an indicator of sympathetic activity [33].

Skin conductivity is monitored using an Empatica E4 wristband. Table 2 [65] shows the characteristics of the sensors that the Empatica E4 device integrates. This device is certified as CE Medical class 2a [66] and has been validated in many works [67,68]. It includes a photo-plethysmography sensor that allows us to measure the blood volume pulse. It also has a galvanic sensor to measure sympathetic nervous system arousal as well as to derive features related to stress, engagemen<sup>t</sup> and excitement. The wristband features are a 3-axis accelerometer to capture motion-based activity and an infrared ray which reads peripheral skin temperature. This device has been designed for continuous, real-time data acquisition.


**Table 2.** Empatica E4 specifications. Data from [65].

However, the calculation of the amplitude is not trivial. Usually, the SCRs overlap each other. In the standard peak detection method (trough-to-peak), the SCR amplitude is obtained by calculating the difference between the peak and the previous trough of the skin conductance data. This results in an underestimation of the amplitude of subsequent SCRs. The degree of underestimation depends on the amplitude and proximity of the preceding SCRs. There are different proposals in the literature to avoid this problem. In this paper, a deconvolution approach [35] is used, which separates skin conductivity data into continuous signals of tonic and phasic activity. This algorithm allows us to represent the overlapping SCRs by compact impulses, thus avoiding the underestimation problem. To that end, we use Ledalab 3.4.9 [69], which is recommended by Empatica. Before the signal deconvolution by continuous decomposition analysis, we pre-process it to eliminate high-frequency noise by applying a smoothening filter consisting of a 4-sample Gaussian window.
