Differential Neurophysiological and Autonomic Responses to Electric, Hybrid, and Internal Combustion Engine Vehicles During Real-World Driving Testing: A Multimodal Psychophysiological Study

Featured Application

The findings of this study have direct applications in automotive engineering, human factors research, and intelligent driver monitoring systems. By demonstrating that vehicle propulsion technology influences drivers cortical activity, autonomic responses, and psychophysiological state, the results provide objective evidence that vehicle cabin environments can affect cognitive workload and physiological stress. These insights may support the next-generation electric and hybrid vehicles optimized for driver comfort and attention. The identified response patterns can contribute to new adaptive driver-assistance systems capable of real-time assessment of workload, fatigue and stress for road safety.

Abstract

The transition from internal combustion engine (ICE) vehicles toward hybrid and battery electric vehicles is reshaping both automotive engineering and the sensory experience of driving, yet its neurophysiological consequences remain underexplored. This single-subject, repeated-measures study investigated the neurophysiological and autonomic responses of one professional driver across 37 real-road sessions in three vehicle categories: electric (EV; n = 16), internal combustion engine (ICE; n = 15), and hybrid (n = 6). Electroencephalography, heart rate, heart-rate variability, galvanic skin response, and peripheral oxygen saturation were continuously recorded across three phases: a three-minute pre-drive baseline, 15–20 min of active driving, and a three-minute post-drive recovery. EEG power was analysed in the theta, alpha, low-beta, and high-beta bands. Because all sessions were completed by the same driver, vehicle group, phase, and their interactions were tested with a rank-based mixed-effects model treating vehicle session as a random intercept. Vehicle group significantly affected broadband EEG power: ICE sessions showed higher theta, low-beta, and high-beta power than EV sessions (all post hoc p ≤ 0.025), with the Hybrid group statistically intermediate; a nonparametric alpha-band effect (p ≤ 0.025) did not survive a repeated-measures-adjusted mixed-effects model (p = 0.107) and is reported with lower confidence. Galvanic skin response declined steeply and uniformly from the pre-drive baseline in every group (phase p < 0.001), whereas heart-rate variability (RMSSD) increased during active driving (p = 0.013). No modality showed a significant Group × Phase interaction (all p ≥ 0.053). Within this single-driver protocol, propulsion type was systematically associated with cortical arousal—ICE imposing a higher cortical load than EV—while autonomic session dynamics were uniform across vehicle types; findings cannot be generalized beyond this driver.

1. Introduction

The global automotive sector is undergoing one of the most consequential technological transitions in its history. Accelerating regulatory pressure, climate commitments, and rapid advances in battery technology are driving an irreversible shift from the internal combustion engine (ICE) paradigm toward battery electric vehicles (EVs) and hybrid electric vehicles (HEVs) [1]. By 2030, EVs and HEVs are projected to constitute the majority of new vehicle sales across major markets, fundamentally reshaping not only transportation infrastructure but also the sensory and psychophysiological experience of mobility itself [2]. Yet while the engineering and environmental dimensions of this transition have been extensively studied, its neurophysiological and human factor implications remain comparatively underexplored [3,4].

Driving is among the most cognitively and physiologically demanding activities of everyday life, requiring the continuous integration of sensory perception, attentional control, motor coordination, spatial navigation, affective regulation, and rapid decision-making under time pressure [5]. The physical environment of the vehicle cabin—its acoustic character, vibrotactile profile, thermal climate, and ergonomic configuration—constitutes a persistent background stimulus that modulates cortical arousal, autonomic nervous system tone, and subjective comfort across the entire driving session [6,7]. ICE, EV, and HEV propulsion technologies differ fundamentally in precisely these environmental dimensions: combustion engines generate continuous broadband noise, mechanical vibration, and exhaust-derived air contaminants, whereas electric drivetrains are substantially quieter, mechanically smoother, and produce no direct cabin emissions [8,9,10]. These differences are not merely a matter of comfort—they represent a systematic manipulation of the physiological environment in which cognition and emotion operate during driving.

Objective assessment of driver physiological state has advanced considerably with the adoption of wearable bio sensing technologies. Electroencephalography (EEG) provides millisecond-resolution access to cortical oscillatory dynamics that index cognitive workload, attentional engagement, vigilance, and fatigue [11,12,13], with frontal-theta dynamics specifically linked to driver fatigue [14,15]. Electrocardiography-derived measures of heart rate (HR) and heart rate variability (HRV) reflect the balance between sympathetic and parasympathetic autonomic regulation, providing sensitive indicators of cardiovascular stress, emotional arousal, and cognitive effort [16,17,18]. Galvanic skin response (GSR), or electrodermal activity, measures sympathetic activation of eccrine sweat glands and provides a continuous, valence-independent index of emotional arousal intensity [4,19]. Pulse oximetry (SpO₂) enables non-invasive monitoring of peripheral oxygen saturation, sensitive to respiratory demand, metabolic rate, and cabin air quality differences that may vary systematically across vehicle types [20,21]. When deployed simultaneously, these signals provide a comprehensive and mechanistically informative window onto the concurrent cognitive, autonomic, and metabolic consequences of driving under differing vehicular conditions—a level of physiological resolution that no single modality can achieve in isolation [3,22,23,24].

Despite the theoretical rationale for expecting substantial physiological differences between ICE, EV, and HEV driving environments, empirical evidence remains limited and methodologically fragmented. The most commonly studied dimension is acoustic: the ICE cabin is substantially louder than the EV cabin, with engine-generated low-frequency noise documented to elevate sympathetic cardiovascular activity, impair parasympathetic recovery, and increase perceived stress during prolonged exposure [25,26]. Comparative biometric studies have reported lower sweat rate, HR, and GSR in EV drivers relative to combustion vehicle equivalents. Cabin air quality—modulated by ventilation mode and combustion-derived contaminants—represents an additional vehicle-type-dependent variable with documented effects on cognitive function and potential SpO₂ dynamics [27,28]. Hybrid vehicles occupy an intermediate and dynamically variable sensory position, with propulsion mode transitions potentially inducing episodic attentional and autonomic responses not present in pure-mode alternatives.

Despite rapid advances in automotive NVH engineering and active road noise control, most studies still evaluate system performance primarily in terms of sound pressure level reduction and objective acoustic metrics, with limited attention to drivers’ neurophysiological and autonomic responses under real world driving conditions. Recent work on road noise control demonstrates effective attenuation of low frequency tire–road noise using feedforward and feedback ANC architectures, yet these evaluations rarely incorporate multimodal psychophysiological endpoints such as EEG, HRV, GSR, and SpO₂ at the driver’s body. No work has systematically tested whether state of the art noise control technologies—passive or active—can normalize or optimize drivers’ neurophysiological profiles in real traffic. This defines a critical research gap: the lack of controlled on road experiments that link specific noise control interventions to changes in drivers’ cortical workload, autonomic regulation, and subjective comfort across different propulsion technologies.

For road noise exposure: Tire–road interaction has become the dominant noise source in modern vehicles, especially at typical urban and highway speeds and in electric vehicles where powertrain noise is reduced. Chronic exposure to low frequency traffic noise is associated with elevated sympathetic activity, impaired parasympathetic recovery, and increased cardiovascular risk, positioning cabin acoustics as a non-trivial determinant of driver health and performance. Cabin sound quality and NVH: Classical NVH work has characterized interior sound quality using psychoacoustic metrics (loudness, sharpness, roughness) and linked them to perceived comfort and brand identity [29,30]. Combustion cabins typically exhibit broadband engine noise and vibration, whereas electric drivetrains yield quieter, smoother environments with different spectral content and masking properties, implying distinct sensory and affective experiences for drivers.

Several works focus on active noise control (ANC) in vehicles. Feedforward and feedback ANC architectures: Active road noise control systems use chassis or cabin sensors to predict or measure primary noise and generate anti-noise via loudspeakers, targeting low frequency components that passive treatments cannot efficiently attenuate. Recent feedback ANC designs based on single microphone error sensing and FxLMS algorithms have demonstrated effective attenuation of measured road noise in simulation and prototype vehicles: “A simulation was implemented based on measured real road noise data, and the simulation results indicate that the proposed feedback ANC system with the single microphone sensor can effectively attenuate road noise.” However, these studies focus on acoustic performance and do not quantify downstream effects on driver physiology or cognition. Several manufacturers report production level RNC systems achieving 3–4 dB overall reductions and larger peak reductions in targeted bands, yet subjective improvements are often modest, suggesting that SPL metrics alone may not capture the full experiential impact of noise control.

The state-of-the-art studies of psychophysiology and neuroergonomics of driving show that neuroergonomics research has established EEG, HRV, GSR, and SpO₂ as sensitive markers of mental workload, vigilance, fatigue, and emotional arousal during driving. Frontal midline theta indexes cognitive workload and executive control, alpha reflects cortical relaxation and attentional engagement, and beta activity tracks sensory arousal and sensorimotor activation. HRV metrics (SDNN, RMSSD) capture sympathovagal balance, while GSR provides a valence-independent index of autonomic arousal intensity. Existing ANC and NVH studies rarely integrate these psychophysiological measures into their evaluation frameworks, and neuroergonomics work has largely treated cabin noise as a background variable rather than a manipulable design parameter. The intersection—systematic testing of noise control technologies with multimodal physiological endpoints in real traffic—remains underdeveloped, motivating our research.

For the road noise control technology and limitations Liu & Lee [31] demonstrate feasibility of feedback ANC for road noise using FxLMS and measured road noise data; they showed effective attenuation but no physiological outcomes, supporting the gap between acoustic and psychophysiological evaluation. Chen et al. [32] review three decades of RNC development, industrial implementations, and remaining challenges in performance, robustness, and system complexity—ideal for framing the technological state of the art and unmet expectations. Regarding the health and neurophysiology of noise exposure, Babisch [25] provides a foundational review linking environmental noise to cardiovascular and autonomic effects. This research supports the claim that cabin noise has health relevant consequences. Münzel et al. [26] details cardiovascular effects of environmental noise exposure, reinforcing the mechanistic plausibility of noise induced autonomic changes in drivers. In terms of psychophysiology of driving and workload, Borghini et al. [3] performed a comprehensive review of EEG, ECG, and EDA for assessing mental workload, fatigue, and drowsiness in pilots and drivers—anchoring the multimodal measurement framework. Jap et al. [33] and Chikhi et al. [34] perform empirical and meta-analytic work, providing evidence that EEG spectral components—especially frontal theta—reliably index cognitive workload during driving and complex tasks. The presented multimodal real-world study in this work provides direct evidence that propulsion type modulates broadband EEG and autonomic responses in real-world driving, establishing the need to test whether noise control interventions can modify these profiles. Collectively, these findings demonstrate that vehicle propulsion technology constitutes a significant and physiologically consequential determinant of driver cortical and autonomic state, with ICE operation imposing a higher cortical load than EV operation and with the Hybrid group occupying an intermediate position.

The present study was designed to provide a comprehensive, multimodal neurophysiological comparison of driving in EV, ICE, and HEV conditions under real-world traffic. A single professional driver completed thirty-seven standardized driving sessions, one per vehicle, with simultaneous recording of EEG (theta, alpha, low beta, high beta), HR, HRV (SDNN and RMSSD), GSR, and SpO₂ throughout a structured three-phase protocol—pre-drive baseline (PREDRIVE), active driving (DRIVE), and post-drive recovery (POSTDRIVE). This design enables three complementary analyses: (1) between-group comparisons of physiological activation as a function of vehicle type; (2) within-session phase analyses examining which modalities are sensitive to the driving episode; and (3) pre-to-post-drive contrasts testing whether the driving session produces a lasting physiological shift that differs between vehicle types. By integrating cortical oscillatory, cardiovascular autonomic, electrodermal, and peripheral oxygenation measures within a single study, we aim to provide a richer account of vehicle-type-dependent driver physiology than any single-modality approach could achieve.

Research Objectives and Hypotheses

Objective 1—Group effects. To determine whether EEG power spectral density, HR, HRV, GSR, and SpO₂ differ significantly between EV, ICE, and HEV driving groups. We hypothesized that ICE drivers would exhibit higher broadband EEG power, elevated HR, reduced HRV, greater GSR, and lower SpO₂ relative to EV drivers, with HEV drivers occupying an intermediate profile.

Objective 2—Phase effects. To characterize within-session temporal dynamics of each physiological measure across PREDRIVE, DRIVE, and POSTDRIVE phases. We hypothesized that active driving would produce selective EEG theta augmentation, HR elevation, HRV reduction, and GSR increase relative to the resting baseline.

Objective 3—PREDRIVE vs. POSTDRIVE comparisons. To test whether the driving session produces a net lasting physiological shift from pre- to post-drive rest, as an index of cumulative fatigue or autonomic residue effects specific to each vehicle type.

Objective 4—Group × Phase interactions. To examine whether the trajectory of physiological change across the session differs between vehicle groups, as evidence that vehicle propulsion type moderates the psychophysiological dynamics of the driving episode.

2. Methodology

2.1. Participants and Experimental Design

A total of 37 vehicle driving sessions were recorded across three vehicle propulsion categories: electric vehicles (EV; n = 16), internal combustion engine vehicles (ICE; n = 15), and hybrid vehicles (Hybrid; n = 6). Each session was conducted by a single professional test driver who operated all 37 vehicles, providing an internally controlled within-driver comparison while capturing the physiological variation attributable to vehicle characteristics. The experimental protocol comprised three sequential recording phases per vehicle: a pre-drive resting baseline (PREDRIVE; 3 min, seated and stationary), an active driving phase (DRIVE; approximately 15–20 min of real-road operation under standardized urban and suburban traffic conditions), and a post-drive recovery period (POSTDRIVE; 3 min, seated and stationary following completion of the drive). All sessions were conducted on the same route under comparable time-of-day and weather conditions to minimize confounding environmental variation. Vehicle testing order was not randomized or counterbalanced with respect to propulsion category; sessions were scheduled pragmatically according to when each vehicle became available for testing (recordings were spread across separate days from March to October 2023), so the testing sequence reflects vehicle availability rather than a predetermined randomization scheme. Driver-state variables that can influence physiological baselines—including prior sleep quality, subjective fatigue, caffeine or alcohol intake, and time since the last meal—were not systematically recorded before each session; the resulting uncontrolled intra-individual variability is considered further in Section 4.7.

Not all sessions yielded valid data in every modality. Note also that the Hyundai NEXO (a hydrogen fuel-cell electric vehicle) is classified as EV throughout, consistent with its all-electric drivetrain; earlier internal processing of the raw per-session physiological data files had mislabeled this vehicle as Hybrid, which was the root cause of an apparent group-size inconsistency in an earlier draft and has now been corrected at the source. The EEG dataset comprised 36 valid sessions (EV = 16, ICE = 14, Hybrid = 6); one ICE session’s EEG recording was unusable due to signal-quality failure and was excluded prior to analysis. The GSR and SpO₂ datasets retained data from all 37 sessions (EV = 16, ICE = 15, Hybrid = 6).

Using a single professional driver does not allow generalization to the population of all drivers. The choice of a single professional driver was intentional and is methodologically justified for the specific research task. The aim of the study is to examine how vehicle propulsion type (EV, HEV, ICE) modulates neurophysiological and autonomic responses, and not to characterize inter individual differences among drivers. Using multiple drivers would introduce substantial uncontrolled variability in driving style, emotional reactivity, baseline physiology, and familiarity with vehicle classes—factors known to strongly influence EEG, HRV, and GSR measures. By holding the driver constant, we ensured that all observed differences arise from the vehicles themselves, not from between driver variability. This approach is consistent with established practice in automotive engineering and NVH research, where a single trained driver is commonly used to ensure repeatability and minimize behavioral confounds. A sample of the fleet that was tested is shown on Figure 1.

Class Electro Vehicle (EV) was represented with (BMW IX, HONDA E NY1, HYUNDAI IONIQ 6, KIA EV9, MERCEDES AMG EQE SUV, MUSTANG MACH E, PEUGEOT E 2008, POLSTAR 2, PORSCHE TAYCAN, SMART BRABUS, SUBARU SOLTERRA, TOYOTA BZ4X, VW E ID5, VW ID BUZZ, VW ID3 FACELIFT, HYUNDAI NEXO).

Class Hybrid was represented with (RENO ARCANA, BMV XM, MERCEDES GLC COUPÉ, MERCEDES S580, VOLVO XC60, VW TOUAREG R EHYBRID).

Class Internal Combustion Engine (ICE) was represented with (ALFA GIULIETTA, BMV X6M, BMW M2, DACIA SANDERO STEPWAY, DACIA JOGER, HAVAL DARGO, HYNDAI BAYON, INEOS GRENADIR, LAND ROVER DEFENDER V8, MERCEDES GLC, FORD MUSTANG MACH1, OPEL ASTRA GSI, VW ARTEON, HYUNDAI STARIA, MERCEDES T DISEL).

2.2. EEG Acquisition and Analysis

Electroencephalography (EEG) is a recording of the brain’s bioelectrical activity. EEG signal detection is done using electrodes, which are placed on the scalp in order to detect electrical activity in the human brain. The neural cells in the human brain communicate by electrical impulses at all instances of time. This activity is monitored in the form of brainwaves. EEG signals can be separated into several waveband classes-based frequency ranges. Electroencephalographic activity was recorded using the EMOTIV EPOC X wireless EEG headset, positioned according to manufacturer guidelines prior to each session with signal quality verified before recording commenced. EEG data were recorded using the EMOTIV EPOC X 14-channel wireless EEG headset. Sensors were positioned according to the international 10–20 system at AF3, F7, F3, FC5, T7, P7, O1, O2, P8, T8, FC6, F4, F8, and AF4, with CMS/DRL (Common Mode Sense/Driven Right Leg) reference sensors at P3/P4. The device applies built-in digital filtering, including a 5th-order Sinc filter, an effective bandwidth of 0.2–45 Hz, and 50/60 Hz notch filtering to attenuate line noise. Band-power values were obtained from the EMOTIV Cortex POW stream, where each value represents absolute band power calculated over the preceding 2 s of EEG data. Artefact-contaminated segments, including those affected by gross movement or poor sensor contact, were excluded before statistical analysis, and additional outlier removal was applied to the derived band-power values. EEG spectral power was extracted and analysed in four canonical frequency bands: theta (4–8 Hz), alpha (8–12 Hz), low beta (β Low, 12–20 Hz), and high beta (β High, 20–30 Hz). Theta power was computed as the average of F3 and F4 electrode positions, reflecting frontal midline theta associated with cognitive workload and executive processing. Alpha power was derived from P7 and P8 (parietal sites), indexing cortical relaxation and attentional engagement. Low and high beta power were computed from the average of F3, F4, P7, and P8 electrode positions, capturing arousal-related and sensorimotor cortical activation.

The electrode placement positions are determined by dividing the skull into perimeters by connecting few reference points on human head. From these points, the skull perimeters are measured in the transverse and median planes. Electrode locations are determined by dividing these perimeters into 10% and 20% intervals. The EMOTIV EPOC X is the 14-channel wireless EEG headset used in this study; data are transmitted to the PC via Bluetooth, sampled internally at 2048 Hz and delivered at a 128 Hz output rate. Based on the cognitive load the cognitive index is presented:

$$\partial =100\left({\displaystyle \frac{\in \mathrm{b}-\in \mathrm{t}}{\in \mathrm{b}}}\right)$$

(1)

where ∂ is the cognitive index, ∈b is the baseline intermission of band power, and ∈t is the task internal of band power. Measurement of cognitive load typically consists of band extraction, feature extraction and classification as shown in Figure 2.

2.3. Cardiovascular, Oximetric, and Electrodermal Measurements

Heart rate (HR; bpm) and peripheral blood oxygen saturation (SpO₂; %) were acquired continuously throughout all three phases from a MAX30100 integrated pulse-oximeter module, with the sensor placed on the right ring finger; the module’s photoplethysmographic (PPG) signal was logged at approximately 100 Hz as an AC-coupled (baseline-removed) waveform, and heart rate, SpO₂, and beat-to-beat R-R intervals were all derived from this same signal. HRV was quantified using two time-domain indices: SDNN (standard deviation of normal-to-normal R-R intervals, reflecting overall HRV) and RMSSD (root mean square of successive R-R differences, reflecting short-term vagal parasympathetic modulation) [35,36]. Electrodermal activity was assessed via galvanic skin response (GSR; Relative Units), with electrodes attached to the index and middle fingers. GSR was recorded continuously across PREDRIVE, DRIVE, and POSTDRIVE phases to quantify sympathetic nervous system activation and psychophysiological arousal.

Skin conductance response (SCR), also referred to as electrodermal activity (EDA) or galvanic skin response (GSR), measures the momentary increase in electrical conductivity of the skin in response to physiologically arousing stimuli, whether external or internal. This phenomenon occurs because emotional activation triggers the sympathetic nervous system, resulting in increased sweat gland activity, which enhances the skin’s electrical conductance properties. When measuring SCR, a mild electrical current is applied across the skin’s surface, allowing researchers to detect fluctuations in conductance levels that correspond to autonomic nervous system activation. These fluctuations directly reflect changes in the activity level of eccrine sweat glands, providing a quantifiable indicator of psychological arousal. It is important to note that arousal represents one fundamental dimension of emotional response, specifically relating to activation intensity rather than emotional valence (positive/negative quality). While not a comprehensive measure of emotion itself, arousal serves as a critical component in understanding emotional states and has demonstrated strong correlations with attentional processes and memory formation. For this study, measurements were obtained using a custom-developed sensor with a 10 Hz sampling rate (approximately 10 samples per second). The device attaches to a participant’s finger and transmits collected data to a computer via USB interface, enabling real-time monitoring of autonomic responses during driving scenarios.

Heart rate, measured in beats per minute (BPM), provides a straightforward and reliable indicator of physiological activation. While breathing rate offers similar insights, heart rate monitoring presents significant advantages in terms of ease of data collection and analysis, making it the preferred metric for assessing autonomic nervous system activity.

Elevated heart rate typically indicates increased arousal levels, though baseline measurements vary between individuals due to factors including body temperature, psychological stress, and sleep quality. Research applications generally focus on heart rate variability (HRV)—the analysis of changes from established baseline values rather than absolute measurements—to detect meaningful physiological responses to stimuli. The primary advantages of heart rate monitoring include its non-invasive nature, cost-effectiveness, and high reliability. Similar to skin conductance response, heart rate measurements reveal how the autonomic nervous system responds to environmental or psychological stimuli, providing valuable insights into a subject’s level of excitement or stress. In this study, heart rate was recorded with the MAX30100 pulse-oximeter module described above (Section 2.3), not with a wrist-worn or smartwatch device.

Blood oxygen saturation (SpO₂) is a critical physiological parameter that measures the percentage of hemoglobin binding sites in the bloodstream occupied by oxygen molecules. In healthy individuals at rest, this value typically ranges between 95 and 100%. This measurement provides valuable insights into respiratory function, cardiovascular response, and autonomic nervous system activity. Modern SpO₂ monitoring utilizes pulse oximetry, a non-invasive technology that employs light-based sensors typically placed on the fingertip or earlobe. These sensors emit specific wavelengths of light that are differentially absorbed by oxygenated and deoxygenated hemoglobin, allowing for real-time monitoring of blood oxygen levels. Contemporary devices are compact, wireless, and capable of continuous monitoring with minimal interference to natural driving behaviors, making them ideal for in-vehicle studies.

The complete sensor and electrode placement configuration used for physiological data acquisition is shown in Figure 3.

2.4. Statistical and Clustering Analysis

EEG power values and physiological variables demonstrated non-normal distributions, confirmed by visual inspection and the pronounced positive skew characteristic of power spectral data. Accordingly, all inferential analyses employed nonparametric statistical procedures. Prior to analysis, outliers in EEG PSD data were identified and removed using a conservative Tukey IQR fence criterion (Q1 − 3 × IQR; Q3 + 3 × IQR) applied independently within each Band × Phase stratum, a more conservative threshold than the standard 1.5 × IQR criterion to avoid excessive data loss in moderately skewed physiological data.

The initial objective is to estimate the appropriate number of clusters for the dataset. Given the nature of the subjects under study—passenger car vehicles—it is reasonable to assume a division into three principal categories: Internal Combustion Engine (ICE) vehicles, Hybrids and Electric Vehicles (EV). Consequently, a natural hypothesis is that the data should exhibit a structure corresponding to at least three distinct clusters. Therefore, applying the k-means clustering algorithm with k = 3 is a logical starting point for uncovering this inherent structure.

Additionally, an alternative theoretical framework supports the use of eight clusters. This framework is based on the conceptual division of the feature space into the eight octants of a cube defined by three binary dimensions: (1) an increase or decrease in the theta/alpha EEG ratio, (2) acceleration or deceleration of heart rate, and (3) an increase or decrease in galvanic skin response. Each of these physiological indicators contributes a dichotomous axis, and their combinations yield eight possible states, each potentially corresponding to a distinct behavioral or cognitive profile. As such, clustering the data into eight groups may also be justified under this psychophysiological model.

The second objective is to evaluate the extent to which the clusters derived from the unsupervised algorithm correspond to known a priori classifications. Specifically, when applying k-means clustering with k = 3, the goal is to assess whether the resulting clusters align with the predefined vehicle categories: ICE, Hybrids and EVs.

Furthermore, when clustering with k = 8 the analysis aims to determine whether the resulting clusters reflect psychophysiological states characterized by combinations of three binary dimensions observed post-driving: (1) increased or decreased heart rate, (2) increased or decreased theta/alpha ratio in EEG activity, and (3) increased or decreased skin conductance level. This evaluation seeks to validate whether the clustering outcomes are consistent with hypothesized physiological response patterns elicited by different vehicle types or driving experiences.

The methodology is that we recorded electrophysiological data from the same driver, who on different days operated various makes of vehicles under urban driving conditions. Data collection was conducted in three distinct phases: (1) a baseline recording lasting 3 min immediately prior to the start of driving; (2) a driving phase recording of variable duration between 15 and 20 min during actual vehicle operation; and (3) a post-driving recording, also lasting 3 min, conducted immediately after the completion of each driving session. All physiological signals—EEG, heart rate, SpO₂, and galvanic skin response—were acquired and pre-processed using the instrumentation and procedures detailed in Section 2.2 and Section 2.3; for the clustering analysis, only the EEG activity from the F3, F4, P7, and P8 positions was used. To this end, the vehicles were initially categorized into three clusters corresponding to their propulsion systems: internal combustion engine (ICE) vehicles, electric vehicles (EVs), and hybrid vehicles.

Additionally, following a baseline normalization procedure, the vehicles were further stratified into eight distinct regions within a three-dimensional feature space defined by changes in three physiological measures recorded pre- and post-driving: electroencephalographic (EEG) theta/alpha ratio, heart rate (pulse rate), and skin conductance level (SCL). These eight regions represent the octants of a conceptual cube, capturing binary increases or decreases along each physiological dimension.

Since some clustering algorithms require the expected number of clusters to be known in advance and effectively “force” the data into one cluster or another, we subjected the data to a test to estimate the expected number of clusters before applying the clustering algorithms.

The elbow method indicated that it makes sense to look for three clusters in the data and to check whether they correspond to the three main types of engines in the studied vehicles, namely internal combustion engine, electric vehicle, and hybrid. On the other hand, the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC) suggest the possibility of seven clusters.

Clustering algorithms description:

$\underset{C_1{, \dots ,C}_k}{\min }{\displaystyle {\sum }_{i=1}^k}\sum _{x{\in C}_i}{\left|x-{\mu }_i\right|}^2{\mu }_i$ K-Means Clustering: it assumes spherical clusters. The approach is fast and scalable. The aim here is to find, where the centroid (mean) of cluster is:

$$C_i: {\displaystyle \frac{1}{\left|C_i\right|}}{\sum }_{x{\in C}_i}x.$$

(2)

Algorithm: Initialization Choose k initial centroids ${\mu }_1^{\left(0\right)}, {\mu }_2^{\left(0\right)}, \dots , {\mu }_k^{\left(0\right)}.$ These can be randomly selected points from the dataset. Assignment Step (E-step) Assign each data point $x_j$ to the nearest centroid:

$${\mathrm{C}}_{\mathrm{i}}^{\left(\mathrm{t}\right)}=\left\{{\mathrm{x}}_{\mathrm{j}}:{\left|{\mathrm{x}}_{\mathrm{j}}-{\mathrm{\mu }}_{\mathrm{i}}^{\left(\mathrm{t}\right)}\right|}^2\le {\left|{\mathrm{x}}_{\mathrm{j}}-{\mathrm{\mu }}_{\mathrm{l}}^{\left(\mathrm{t}\right)}\right|}^2⩝\mathrm{l}=1,\dots ,\mathrm{k}\right\}$$

(3)

Update Step (M-step) Update each centroid to be the mean of the points assigned to it:

$${\mu }_i^{(t+1)}={\displaystyle \frac{1}{\left|C_i^{\left(t\right)}\right|}}{\sum }_{x_j\in C_i^{\left(t\right)}}{x}_j$$

(4)

Steps 2–3 are repeated until convergence (i.e., centroids no longer change significantly or assignments stop changing).

Convergence Criterion: Typically, the algorithm stops when: ${\displaystyle {\sum }_{i=1}^k}{\left|{\mu }_i^{(t+1)}-{\mu }_i^{\left(t\right)}\right|}^2<\epsilon$, for small $\epsilon > 0$, or a fixed number of iterations is reached.

Determining the optimal number of clusters (denoted as k) in k-means clustering is a crucial step, as it directly affects the quality of the clustering. Here are the most commonly used methods to determine the best value for k: 1. Elbow Method: Concept: Plot the Within-Cluster Sum of Squares (WCSS) against different values of k. WCSS measures the variance within each cluster. To calculate this, we need to do the following: Compute k-means clustering for a range of k (e.g., from 1 to 10). Plot the number of clusters vs. WCSS. Look for an “elbow” point where the rate of decrease sharply shifts. This point suggests diminishing returns for adding more clusters. Other methods are Silhouette Score. Gap Statistic, Calinski–Harabasz Index, Davies–Bouldin Index.

DBSCAN: Finds arbitrarily shaped clusters, Can detect noise/outliers.

DBSCAN groups together points that are closely packed, and marks as outliers the points that lie alone in low-density regions. Input Parameters: ε > 0: radius (neighborhood size), MinPts ∈ N: minimum number of points required to form a dense region.

Definitions: Let X = $\left\{x_1, x_2,\dots ,x_n\right\}\subset R^d$ be the dataset; $\epsilon$-neighborhood of a point x: $N_{\epsilon }\left(x\right)=\left\{y\in \mathrm{X}|\left|x-y\right|\le \epsilon \right\}$, core point: A point x is a core point if:

$\left|N_{\epsilon }\left(x\right) \ge MinPts\right| \ge MinPts$; directly density-reachable: A point y is directly density-reachable from x if: x is a core point and $y\in N_{\epsilon }\left(x\right)$; density-reachable: A point y is density-reachable from x if there exists a sequence: $x = x_0 , x_1 , \dots , x_k = y$, such that $x_{i+1}$, is directly density-reachable from $x_i$ for all i, and $x_0$ is a core point; density-connected: Two points x and y are density-connected if there exists a point z such that both x and y are density-reachable from z.

For each unvisited point $x\in X$: Mark x as visited. If x is a core point, start a new cluster C and expand it: Add all points that are density-reachable from x to C. If x is not a core point and not density-reachable from any core point, mark it as noise.

Gaussian Mixture Models (GMM): Assumes clusters follow Gaussian distributions.

A Gaussian Mixture Model (GMM) is a probabilistic model that assumes that all the data points are generated from a mixture of several Gaussian distributions with unknown parameters. It is commonly used for clustering, density estimation, and unsupervised learning. Each Gaussian distribution in the mixture is defined by: A mean vector (μ), A covariance matrix (Σ), A mixing coefficient (π) representing the weight or proportion of each component. GMM is often learned via the Expectation-Maximization (EM) algorithm. Algorithm: Expectation-Maximization (EM) for GMM. Input: Data:

X = $\left\{x_1,x_2,\dots ,x_N\right\}$, where $x_i\in R^d$, Number of components: K, Convergence threshold: $\epsilon$ Estimated parameters: $\left\{{\pi }_k,{\mu }_k,{\mathsf{\Sigma }}_k\right\}$ for k = 1, …, K.

Step-by-Step Algorithm:

1.

Initialization: Randomly initialize the parameters: Mixing coefficients ${\pi }_k= {\displaystyle \frac{1}{K}}$, Means ${\mu }_k$, Covariances ${\mathsf{\Sigma }}_k$.

2.

Expectation Step (E-step): Calculate the responsibilities (posterior probabilities): ${\gamma }_{ik}={\displaystyle \frac{{\pi }_k.{\rm N}\left(x_i\right|{\mu }_k,{\Sigma }_k)}{{\displaystyle {\sum }_{j=1}^K}{\pi }_{j.}{\rm N}(x_i|{\mu }_j,{\Sigma }_j)}}$, where ($x_i|{\mu }_k,{\mathsf{\Sigma }}_k$) is the Gaussian distribution with mean ${\mu }_k$ and covariance ${\mathsf{\Sigma }}_k$.

3.

Maximization Step (M-step) Update parameters using the responsibilities:

$$N_k={\displaystyle {\sum }_{i=1}^N}{\gamma }_{ik};{\pi }_k={\displaystyle \frac{N_k}{N}};{\mu }_k={\displaystyle \frac{1}{N_k}}{\displaystyle {\sum }_{i=1}^N}{\gamma }_{ik}{x}_i;{\sum }_k={\displaystyle \frac{1}{N_k}}{\displaystyle {\sum }_{i=1}^N}{\gamma }_{ik}(x_i-{\mu }_k){(x_i-{\mu }_k)}^T$$

4.

Convergence Check: Calculate the log-likelihood:

$$\log L={\displaystyle {\sum }_{i=1}^N}\log ({\displaystyle {\sum }_{k=1}^K}{\pi }_k.N(x_i|{\mu }_k,{\mathsf{\Sigma }}_k))$$

(5)

If the change in log-likelihood is less than ϵ\epsilonϵ, stop. Else, go to Step 2.

Agglomerative clustering is a hierarchical clustering method that builds nested clusters by merging or agglomerating data points based on a similarity or distance metric. The process is bottom-up: each data point starts as its own cluster, and pairs of clusters are merged iteratively.

Let $X=\left\{x_1,x_2,\dots ,x_n\right\}\subset R^d$, be a set of n data points in d-dimensional space. Initially, each point is its own cluster: $C^{\left(0\right)}=\left\{\left\{x_1\right\},\left\{x_2\right\},\dots ,\left\{x_n\right\}\right\}$. Define Euclidean distance: ${d(x_i,x_j)=\left|x_i-x_j\right|}^2$ with distance metric $R^d \times R^d\to R$. The distance between clusters is defined as $D(A,B)=\underset{x\in A,y\in B}{\min }d(x,y)$. On each iteration t, the pair of clusters $A,B\in C^{\left(t\right)}$ are merged with smallest inter-cluster distance as: $(A^{*},B^{*})=\underset{A\ne B\in C^{\left(t\right)}}{\min }D(A,B)$. The cluster set is updated according to: $C^{(t+1)}=(C^{\left(t\right)}\backslash \left\{A^{*},B^{*}\right\})\cup \left\{A^{*}\cup B^{*}\right\}$. The algorithm stops when the desired number of clusters k is reached: $\left|C^k\right|=k$.

Clustering implementation and parameter settings. All clustering analyses in this study were implemented in Python 3 using the scikit-learn library (scikit-learn version 1.6.1). For each of the 37 sessions, the clustering input was the post-minus-pre-drive difference (POST − PRE) in four physiological features—the EEG theta/alpha ratio, pulse rate, skin-conductance level, and peripheral oxygen saturation—which were standardized to zero mean and unit variance (z-score standardization) before clustering; the same standardized feature matrix was supplied to all four algorithms. K-Means was applied with the target cluster counts k = 3 and k = 8 using k-means++ initialization, scikit-learn’s default number of initializations, and a fixed random seed (random_state = 42). The Gaussian Mixture Models used 3 and 8 components with full covariance matrices, a single initialization, and the same fixed seed. Agglomerative clustering used Ward linkage with the Euclidean distance metric; the single-linkage expression given above is a generic illustration of the agglomerative scheme rather than the criterion applied here. DBSCAN was applied to the same standardized features, with its neighbourhood parameters tuned by k-distance inspection because DBSCAN does not take a target cluster count: the three-group solution used ε = 0.75 with min_samples = 3 (two dense clusters plus points labelled as noise), whereas the eight-group solution used ε = 1.42 with min_samples = 1, which produced a single dense cluster together with several isolated single-session points rather than eight balanced clusters—accounting for DBSCAN’s low agreement with the other three methods. The eight-cluster reference partition against which the methods were compared is the psychophysiological octant model defined by the three binary dimensions (increase/decrease in the EEG theta/alpha ratio, pulse rate, and skin conductance).

Group differences among EV, ICE, and Hybrid conditions were evaluated using the Kruskal–Wallis H test. Pairwise post hoc comparisons following significant omnibus tests used the Mann–Whitney U test with Bonferroni correction (adjusted α = 0.017 for three pairwise contrasts). Repeated-measures phase effects were analysed using the Friedman test for within-subject comparisons, and the Kruskal–Wallis test for across-subject phase comparisons. Pre-to-post-drive changes (PREDRIVE vs. POSTDRIVE) were assessed with the Wilcoxon signed-rank test. Statistical significance was set at p < 0.05 for all omnibus tests. All analyses were conducted in Python 3 using SciPy and pandas. Because all sessions were completed by the same driver and the three within-session phases (PREDRIVE, DRIVE, POSTDRIVE) are repeated measurements on the same session, observations are not independent and a standard between-groups nonparametric design is inappropriate. Group (vehicle type), phase, and their interaction were therefore additionally tested using a rank-based mixed-effects model (values rank-transformed within each variable, fit with vehicle session as a random intercept and vehicle group × phase as fixed effects), as appropriate for the non-normal, repeated-measures structure of the data; terms were evaluated via likelihood-ratio comparison against nested reduced models. Unless otherwise noted, error bars in all figures represent the interquartile range (Q1–Q3) around the plotted median value, consistent with the descriptive statistics reported in the corresponding descriptive-statistics tables presented in this section. This approach formally tests the vehicle group × phase interaction that the original nonparametric between-groups tests could not address.

3. Results

3.1. EEG Power Spectral Density

3.1.1. Data Preparation and Outlier Removal

The EEG dataset comprised 36 valid vehicle sessions—EV (n = 16), ICE (n = 14), Hybrid (n = 6)—across four frequency bands and three phases, yielding 404 observations before cleaning. A total of 60 data points (14.9%) were identified as extreme outliers and excluded, predominantly corresponding to EMG contamination and electrode artefacts with values exceeding 20–4100 µV²/Hz in the alpha and beta bands. The cleaned dataset retained 344 valid observations. All inferential tests used Bonferroni-corrected nonparametric procedures with α = 0.05.

3.1.2. Descriptive Statistics

Median EEG power spectral density values across all Group × Band × Phase cells are presented in

Table 1. Median (IQR) EEG power spectral density (µV²/Hz) by Vehicle Group, Frequency Band, and Recording Phase (outlier-cleaned data).

Group	Band	PREDRIVE	DRIVE	POSTDRIVE
EV	θ	2.35 (1.92–3.09)	3.81 (3.58–4.01)	2.82 (2.25–3.08)
EV	α	0.81 (0.62–0.99)	0.94 (0.81–1.41)	0.76 (0.67–1.33)
EV	β Low	0.66 (0.56–0.83)	0.77 (0.67–0.91)	0.64 (0.56–0.68)
EV	β High	0.33 (0.28–0.46)	0.40 (0.33–0.47)	0.35 (0.33–0.42)
Hybrid	θ	2.31 (2.23–3.81)	5.37 (5.13–5.61)	7.26 (4.64–7.65)
Hybrid	α	0.71 (0.39–0.83)	0.91 (0.68–4.19)	0.80 (0.50–2.21)
Hybrid	β Low	0.49 (0.45–0.60)	0.66 (0.58–2.68)	0.64 (0.60–1.48)
Hybrid	β High	0.27 (0.25–0.31)	0.34 (0.31–1.26)	0.31 (0.31–0.68)
ICE	θ	3.91 (3.26–5.40)	4.54 (4.07–5.46)	3.70 (2.14–4.10)
ICE	α	2.07 (1.07–2.42)	1.17 (0.85–2.29)	0.95 (0.69–2.12)
ICE	β Low	1.12 (0.88–1.23)	1.00 (0.75–1.22)	0.89 (0.62–1.40)
ICE	β High	0.59 (0.54–0.79)	0.69 (0.40–0.94)	0.44 (0.40–0.83)

Note. θ = theta; α = alpha; β Low = low beta; β High = high beta. IQR = interquartile range (Q1–Q3).

. Across all frequency bands, the ICE group exhibited consistently higher median PSD than the EV group; comparisons with the Hybrid group were less consistent, as Hybrid’s theta median numerically exceeded ICE’s during DRIVE and POSTDRIVE, most likely reflecting the instability of medians computed from the very small Hybrid theta sample (n = 2–3 per cell after outlier removal) rather than a genuine reversal of the group effect (see Section 4.7). The spectral ordering θ > α > βL > βH was preserved in all groups, consistent with the 1/f spectral slope of EEG.

3.1.3. Main Effect of Vehicle Group

Kruskal–Wallis tests revealed a statistically significant main effect of vehicle group on EEG PSD across all four frequency bands (

Table 2. Kruskal–Wallis test results for the main effect of Vehicle Group on EEG PSD.

Band	H Statistic	df	p Value
θ	12.35	2	0.002
α	7.51	2	0.023
β Low	12.79	2	0.002
β High	18.33	2	<0.001

Note. Bonferroni-corrected pairwise contrasts: ICE > EV in all four bands (p_adj ≤ 0.025; but see Section 3.1.6—the alpha effect does not survive the repeated-measures-adjusted model); ICE vs. Hybrid and EV vs. Hybrid non-significant in all bands after correction (all p_adj ≥ 0.070).

). Effect magnitudes increased progressively from alpha through high beta. As detailed in Section 3.1.6, the alpha-band effect does not survive the repeated-measures-adjusted mixed-effects model and should be interpreted with more caution than the theta, low-beta, and high-beta effects.

Post hoc Mann–Whitney U comparisons (Bonferroni-corrected, α_adj = 0.017) consistently identified the ICE group as the primary source of group differences. ICE sessions exhibited significantly higher PSD than EV sessions in all four bands (θ: p_adj = 0.002; α: p_adj = 0.025; βL: p_adj = 0.002; βH: p_adj < 0.001). With the corrected vehicle classification, ICE–Hybrid differences do not survive Bonferroni correction in any band (θ: p_adj = 1.000; α: p_adj = 0.399; βL: p_adj = 0.172; βH: p_adj = 0.070). No significant EV–Hybrid differences were observed in any band (all p_adj ≥ 0.247). When data were pooled across bands on the log-transformed scale, the omnibus group effect remained highly significant (H = 15.08, df = 2, p < 0.001), with ICE > EV confirmed (p_adj < 0.001) and EV vs. Hybrid non-significant (p_adj = 1.000); with the corrected vehicle classification, the ICE > Hybrid contrast reported in an earlier draft (H = 16.40, p_adj = 0.035) no longer survives Bonferroni correction (p_adj = 0.196), consistent with the absence of a significant ICE-Hybrid difference in every individual band reported above.

3.1.4. Main Effect of Recording Phase

Kruskal–Wallis tests for the main effect of recording phase revealed a significant effect exclusively in the theta band (

Table 3. Kruskal–Wallis test results for the main effect of Recording Phase on EEG PSD.

Band	H Statistic	df	p Value
θ	12.82	2	0.002
α	1.33	2	0.515
β Low	2.71	2	0.258
β High	1.39	2	0.499

); alpha and both beta bands were non-significant. A significant low-beta phase effect did, however, emerge under the repeated-measures-adjusted mixed-effects model (Section 3.1.6), indicating that the simpler between-groups test lacked power to detect it.

For theta, DRIVE-phase power was significantly elevated relative to both PREDRIVE (median DRIVE = 4.07 vs. PREDRIVE = 3.08 µV²/Hz; p_adj = 0.015) and POSTDRIVE (median POSTDRIVE = 2.85 µV²/Hz; p_adj = 0.002), while PREDRIVE and POSTDRIVE did not differ (p_adj = 1.000). This theta augmentation during active driving is consistent with established findings linking frontal-midline theta to heightened cognitive workload and executive control during driving [33,37]. The absence of phase effects in alpha and beta bands indicates that power modulation in these ranges was governed primarily by vehicle group membership rather than the temporal task structure.

As a robustness check accounting for the paired, repeated-measures structure of the phase factor, a Friedman test restricted to sessions with complete PREDRIVE–DRIVE–POSTDRIVE triads confirmed a significant theta phase effect (χ²(2) = 14.00, p < 0.001, n = 25, Kendall’s W = 0.28) and additionally revealed a significant high-beta phase effect (χ²(2) = 6.33, p = 0.042, n = 24, Kendall’s W = 0.13). Alpha (χ²(2) = 4.56, p = 0.102) and low beta (χ²(2) = 3.25, p = 0.197) remained non-significant under the paired test.

3.1.5. PREDRIVE vs. POSTDRIVE and Within-Group Phase Effects

No significant PREDRIVE –POSTDRIVE change was detected in any EEG frequency band (θ: W = 146, p = 0.468; α: W = 134, p = 0.458; βL: W = 126, p = 0.509; βH: W = 141, p = 0.812). Within-group Friedman tests (

Table 4. Friedman test results for the effect of Recording Phase within each Vehicle Group and EEG Frequency Band.

Band	Group	χ2(2)	p Value	n
θ	ICE	6.22	0.045	9
θ	EV	10.86	0.004	14
α	ICE	2.00	0.368	9
α	EV	4.00	0.135	14
β Low	ICE	0.67	0.717	9
β Low	EV	5.54	0.063	13
β High	ICE	1.56	0.459	9
β High	EV	6.00	0.050	13

Note. Friedman test could not be performed for the Hybrid group (n ≤ 3 complete triads per cell). Significant results in bold.

) revealed significant phase effects for theta in both ICE (χ²(2) = 6.22, p = 0.045) and EV (χ²(2) = 10.86, p = 0.004) groups, confirming task-coupled theta modulation. The EV group also showed a phase effect in high beta that narrowly survived the 0.05 threshold (χ²(2) = 6.00, p = 0.050), reflecting a DRIVE-phase elevation with post-drive return—a pattern absent in ICE drivers. Low beta remained a non-significant trend in EV (χ²(2) = 5.54, p = 0.063); this should be treated as a non-significant trend rather than a confirmed effect.

The corresponding EEG group-level distributions by Recording Phase and by Vehicle Type, for the theta/alpha and low-beta/high-beta bands, are illustrated in Figure 4, Figure 5, Figure 6 and Figure 7.

3.1.6. Group × Phase Interaction

A rank-based mixed-effects model with vehicle session as a random intercept and vehicle group × phase as fixed effects (Section 2.4) formally tested the Group × Phase interaction while accounting for the repeated within-session structure of the data. No EEG band showed a significant interaction: theta (χ²(4) = 5.28, p = 0.260), alpha (χ²(4) = 6.97, p = 0.138), low beta (χ²(4) = 4.68, p = 0.321), and high beta (χ²(4) = 2.57, p = 0.633). Evaluating main effects by likelihood-ratio tests against nested reduced models, the same model confirmed a significant Group main effect for theta (χ²(2) = 9.54, p = 0.008), low beta (χ²(2) = 7.62, p = 0.022), and high beta (χ²(2) = 10.29, p = 0.006), with alpha non-significant (χ²(2) = 4.47, p = 0.107); and a significant Phase main effect for theta (χ²(2) = 21.32, p < 0.001) and low beta (χ²(2) = 6.32, p = 0.042). These repeated-measures-adjusted results diverge from the nonparametric analyses reported above (

Table 5. Group × Phase interaction and repeated-measures-adjusted main effects (rank-based mixed-effects likelihood-ratio tests), EEG.

Band	Group Main Effect χ2(2), p	Phase Main Effect χ2(2), p	Group × Phase χ2(4), p
Theta	9.54, p = 0.008	21.32, p < 0.001	5.28, p = 0.260
Alpha	4.47, p = 0.107	4.57, p = 0.102	6.97, p = 0.138
Low beta	7.62, p = 0.022	6.32, p = 0.042	4.68, p = 0.321
High beta	10.29, p = 0.006	3.38, p = 0.184	2.57, p = 0.633

) in two respects. First, the alpha-band Group effect, significant under the Kruskal–Wallis and post hoc tests (Table 2; Section 3.1.3), does not survive adjustment for the repeated within-session structure; because the mixed-effects model is the analytically more appropriate test for this non-independent, single-driver design, the alpha finding should be regarded as the weakest and least robust of the four Group effects, and claims of an effect in “all four bands” elsewhere in this paper apply most confidently to theta, low beta, and high beta. Second, the low-beta phase effect, non-significant under the Kruskal–Wallis test (Table 3), reaches significance under the mixed-effects model, suggesting the simpler between-groups test was underpowered to detect it. Both discrepancies most likely reflect the greater sensitivity of the repeated-measures-adjusted model to this dataset’s non-independent structure, rather than a data error.

3.2. Galvanic Skin Response and Peripheral Oxygen Saturation

3.2.1. Descriptive Statistics

GSR and SpO₂ descriptive statistics for all Group × Phase cells are presented in

Table 6. Median (IQR), Mean ± SD for GSR (Relative Units) and SpO₂ (%) by Vehicle Group and Recording Phase.

Measure	Group	Phase	n	Median	Q1	Q3	IQR	Mean	SD
GSR	EV	PREDRIVE	16	443.50	372.81	490.48	117.67	434.86	92.10
GSR	EV	DRIVE	16	301.63	273.01	379.60	106.58	328.71	122.94
GSR	EV	POSTDRIVE	16	276.00	216.08	328.10	112.03	284.67	134.10
GSR	ICE	PREDRIVE	15	458.56	385.74	509.58	123.84	444.29	80.52
GSR	ICE	DRIVE	15	319.49	261.16	400.20	139.04	324.83	111.50
GSR	ICE	POSTDRIVE	15	269.00	238.73	333.84	95.11	276.60	98.18
GSR	Hybrid	PREDRIVE	6	439.44	430.77	487.65	56.88	455.97	44.65
GSR	Hybrid	DRIVE	6	327.05	289.12	368.75	79.63	332.24	87.02
GSR	Hybrid	POSTDRIVE	6	284.94	247.13	299.46	52.33	287.13	88.81
SpO2	EV	PREDRIVE	15	96.00	95.81	96.16	0.35	96.28	1.14
SpO2	EV	DRIVE	14	96.09	95.99	96.29	0.30	96.42	1.08
SpO2	EV	POSTDRIVE	15	96.15	95.86	96.28	0.42	96.36	1.09
SpO2	ICE	PREDRIVE	13	95.97	95.83	96.02	0.19	96.26	1.14
SpO2	ICE	DRIVE	12	95.90	95.58	96.03	0.45	95.69	0.73
SpO2	ICE	POSTDRIVE	13	96.00	95.92	96.07	0.15	95.86	0.61
SpO2	Hybrid	PREDRIVE	6	95.97	95.89	96.47	0.58	96.71	1.64
SpO2	Hybrid	DRIVE	6	96.15	95.97	96.70	0.73	96.82	1.60
SpO2	Hybrid	POSTDRIVE	6	96.31	95.81	96.67	0.86	96.79	1.63

Note. IQR = Q3 − Q1. GSR values in Relative Units; SpO₂ values in %.

. GSR showed a pronounced and consistent decline from PREDRIVE to DRIVE to POSTDRIVE in all three vehicle groups, with median reductions of 35–41% from baseline to post-drive. SpO₂ showed minimal absolute variation across all cells, with median values contained within a narrow range of 95.90–96.31%.

The corresponding distributions of galvanic skin response and peripheral oxygen saturation by Recording Phase and by Vehicle Type are shown in Figure 8 and Figure 9.

3.2.2. Main Effects and PREDRIVE vs. POSTDRIVE

Vehicle group had no significant main effect on either GSR (H(2) = 0.09, p = 0.955) or SpO₂ (H(2) = 4.98, p = 0.083), with all Bonferroni-corrected pairwise contrasts non-significant (all p_adj ≥ 1.000). In striking contrast, recording phase exerted a highly significant and large main effect on GSR (H(2) = 35.90, p < 0.001)—the largest effect in the entire study—characterized by a monotonic decline from PREDRIVE (median = 442.1) through DRIVE (309.8) to POSTDRIVE (280.5), with PREDRIVE significantly exceeding both DRIVE (p_adj < 0.001) and POSTDRIVE (p_adj < 0.001). Recording phase had no significant effect on SpO₂ (H(2) = 0.60, p = 0.741). The PREDRIVE–POSTDRIVE Wilcoxon test confirmed a highly significant GSR reduction in the full sample (W = 8.0, p < 0.001, median Δ = −164.0, ~37% decline) that was significant within the EV (p_adj < 0.001) and ICE (p_adj < 0.001) groups individually; the Hybrid group showed the same directional decline and, at the corrected n = 6. This result (p_adj = 0.094, raw p = 0.031) should still be treated cautiously given the small sample size (Section 4.7). No significant PREDRIVE–POSTDRIVE SpO₂ change was detected at the whole-sample (p = 0.187) or group level.

3.2.3. Within-Group Phase Effects and Interactions

Friedman tests revealed a strong, highly significant phase effect on GSR in the EV (χ²(2) = 17.38, p < 0.001, n = 16), ICE (χ²(2) = 24.40, p < 0.001, n = 15), and Hybrid (χ²(2) = 12.00, p = 0.002, n = 6) groups—indicating that PREDRIVE > DRIVE > POSTDRIVE was the rank order in nearly every session. This universality establishes the GSR session trajectory as a protocol-driven autonomic signature independent of vehicle type, consistent with progressive habituation of anticipatory sympathetic arousal from the pre-drive period through driving and recovery.

For SpO₂, within-group Friedman tests found no significant phase effect in any vehicle group once the corrected vehicle classification is used: EV (χ²(2) = 2.63, p = 0.269, n = 14), ICE (χ²(2) = 3.96, p = 0.138, n = 12), and Hybrid (χ²(2) = 2.80, p = 0.247, n = 6) all showed flat, non-significant SpO₂ trajectories across the session—this differs from an earlier analysis that reported a significant ICE phase effect and a borderline Hybrid effect, both of which were artefacts of the pre-correction vehicle classification. GSR and SpO₂ are treated throughout as independent physiological indicators, and no systematic relationship between them was observed across the session.

3.2.4. Group × Phase Interaction

The same rank-based mixed-effects model (Section 2.4) found no significant Group × Phase interaction for GSR (χ²(4) = 1.26, p = 0.869) or SpO₂ (χ²(4) = 6.43, p = 0.169). It confirmed no significant Group main effect for either measure (GSR χ²(2) = 0.06, p = 0.972; SpO₂ χ²(2) = 2.47, p = 0.290) and a very strong Phase main effect for GSR (χ²(2) = 95.85, p < 0.001) but not SpO₂ (χ²(2) = 1.48, p = 0.478), consistent with the Kruskal–Wallis and Friedman results reported above.

3.3. Heart Rate and Heart Rate Variability

3.3.1. Descriptive Statistics

HR, HRV-SDNN, and HRV-RMSSD descriptive statistics are presented in

Table 7. Median (IQR) and Mean ± SD for Heart Rate (bpm), HRV-SDNN (ms), and HRV-RMSSD (ms) by Vehicle Group and Recording Phase.

Measure	Group	Phase	n	Median	Q1	Q3	Mean	SD
HR	EV	PREDRIVE	11	75.4	74.2	82.2	77.9	4.7
HR	EV	DRIVE	14	76.5	75.3	81.1	77.5	6.5
HR	EV	POSTDRIVE	12	78.5	76.2	79.9	78.2	4.0
HR	ICE	PREDRIVE	12	81.3	76.6	86.3	81.6	6.5
HR	ICE	DRIVE	11	78.7	76.8	84.0	80.6	5.3
HR	ICE	POSTDRIVE	12	81.2	78.6	84.2	81.6	4.8
HR	Hybrid	PREDRIVE	3	76.5	73.9	79.9	77.0	6.0
HR	Hybrid	DRIVE	4	76.3	72.8	80.6	77.0	5.9
HR	Hybrid	POSTDRIVE	3	78.4	77.7	84.2	81.8	7.2
SDNN	EV	PREDRIVE	11	38.5	29.0	50.6	40.8	14.1
SDNN	EV	DRIVE	14	45.2	36.6	47.4	41.6	14.4
SDNN	EV	POSTDRIVE	12	35.0	30.2	40.5	36.3	7.8
SDNN	ICE	PREDRIVE	12	34.4	30.0	41.3	35.9	7.7
SDNN	ICE	DRIVE	11	40.2	35.2	45.3	40.5	6.2
SDNN	ICE	POSTDRIVE	12	43.3	36.5	45.8	42.5	10.5
SDNN	Hybrid	PREDRIVE	3	38.4	31.8	43.4	37.3	11.7
SDNN	Hybrid	DRIVE	4	46.9	38.3	56.3	47.7	17.6
SDNN	Hybrid	POSTDRIVE	3	39.0	32.0	43.0	37.0	11.2
RMSSD	EV	PREDRIVE	11	40.9	36.5	61.0	49.6	21.5
RMSSD	EV	DRIVE	14	54.7	44.6	62.7	52.1	18.5
RMSSD	EV	POSTDRIVE	12	42.5	40.4	49.9	45.4	9.8
RMSSD	ICE	PREDRIVE	12	38.4	32.7	47.9	40.5	10.6
RMSSD	ICE	DRIVE	11	49.2	46.7	57.5	50.7	8.5
RMSSD	ICE	POSTDRIVE	12	46.3	42.2	51.1	45.7	8.0
RMSSD	Hybrid	PREDRIVE	3	39.7	31.8	51.3	42.2	19.5
RMSSD	Hybrid	DRIVE	4	54.1	42.6	65.0	53.5	18.9
RMSSD	Hybrid	POSTDRIVE	3	53.6	45.1	60.0	52.2	15.0

. Missing values were present in 13–31% of EV/ICE cells and 33–50% of Hybrid cells due to recording artefacts, reflecting the small Hybrid enrollment (n = 3–4 effective triads). ICE drivers had numerically the highest HR across all phases; HRV indices were broadly comparable across groups.

3.3.2. Group Effects

Vehicle group had a borderline significant effect on HR (H(2) = 6.08, p = 0.048), with ICE drivers numerically highest (pooled median 80.1 bpm) relative to EV (76.5 bpm) and Hybrid (77.7 bpm); however, no pairwise contrast survived Bonferroni correction (ICE vs. EV: p_adj = 0.052). Vehicle group had no significant effect on either HRV-SDNN (H(2) = 0.05, p = 0.977) or HRV-RMSSD (H(2) = 0.55, p = 0.760), with pooled-phase medians closely matched across groups for both indices.

3.3.3. Phase Effects

Phase had no significant main effect on HR (H(2) = 0.70, p = 0.705)—heart rate was virtually identical across PREDRIVE (78.7 bpm), DRIVE (78.3 bpm), and POSTDRIVE (79.3 bpm). Phase had no significant effect on HRV-SDNN (H(2) = 3.47, p = 0.176), though a numerical trend toward DRIVE-phase elevation (SDNN: PREDRIVE 36.5 → DRIVE 44.0 → POSTDRIVE 39.0 ms) approached but did not reach significance. Phase had a significant effect on HRV-RMSSD (H(2) = 8.66, p = 0.013), driven by a robust PREDRIVE < DRIVE contrast (p_adj = 0.028; pooled medians PREDRIVE = 38.9 ms, DRIVE = 52.0 ms), representing a median DRIVE-phase increase of approximately 34% in vagal cardiac modulation relative to the pre-drive baseline. This finding runs counter to a simple stress-arousal model of driving and is discussed in Section 4.4.

3.3.4. Within-Group Phase Effects

Friedman tests within vehicle groups revealed no significant phase effect on HR in either EV (χ²(2) = 2.91, p = 0.234, n = 11) or ICE (χ²(2) = 0.73, p = 0.695, n = 11) groups, consistent with the null omnibus finding. For HRV-SDNN, both EV (χ²(2) = 5.64, p = 0.060) and ICE (χ²(2) = 5.09, p = 0.078) groups showed trends that approached but did not reach significance. For HRV-RMSSD, significant within-subject phase effects were found in both EV (χ²(2) = 7.80, p = 0.020, n = 10) and ICE (χ²(2) = 10.36, p = 0.006, n = 11) groups, confirming the DRIVE-phase RMSSD elevation as a consistent within-person phenomenon across both vehicle types.

3.3.5. HR–HRV Correlations

Spearman rank correlations between HR and HRV indices revealed theoretically expected negative associations. The HR–SDNN correlation was strongly negative at PREDRIVE (rs = −0.837, p < 0.001), attenuating during DRIVE (rs = −0.488, p = 0.007) and POSTDRIVE (rs = −0.462, p = 0.015). The HR–RMSSD correlation was negligible at PREDRIVE (rs = +0.003, p = 0.988) but became significantly negative during DRIVE (rs = −0.415, p = 0.025), indicating that task-induced sympathetic activation specifically during driving was associated with reduced vagal cardiac modulation.

Heart rate and HRV-RMSSD distributions by Recording Phase are illustrated in Figure 10.

3.3.6. Group × Phase Interaction

The same rank-based mixed-effects model (Section 2.4) found no significant Group × Phase interaction for HR (χ²(4) = 7.40, p = 0.116), HRV-SDNN (χ²(4) = 9.36, p = 0.053), or HRV-RMSSD (χ²(4) = 4.16, p = 0.384). No significant Group main effect was found for any cardiovascular measure (HR χ²(2) = 3.99, p = 0.136; SDNN χ²(2) = 0.21, p = 0.902; RMSSD χ²(2) = 0.52, p = 0.771). A significant Phase main effect was confirmed for RMSSD (χ²(2) = 11.98, p = 0.003), with SDNN a non-significant trend (χ²(2) = 5.28, p = 0.071) and no effect for HR (χ²(2) = 3.00, p = 0.223)—consistent with the DRIVE-phase RMSSD elevation reported above.

The corresponding distributions of HR and HRV-SDNN, and of HRV-RMSSD, by Vehicle Group are presented in Figure 11 and Figure 12.

3.4. Clustering Results

In order not to advertise a particular car model, the individual cars and their presence in one or another cluster will be discussed only under their index number from 1 to 37, which does not correspond to their alphabetical order. The following clustering and multidimensional-scaling analyses are exploratory and hypothesis-generating. With three a priori categories and only 37 objects, the choice of k = 3 and k = 8 clusters is partly determined by the categories and dimensions the analysis was designed to test; these results should not be interpreted as confirmatory.

Clustering into 3 clusters:

Engine Type known clusters: EV (3,4,9,10,11,12,14,19,21,22,27,29,30,31,33,37), ICE (1,2,5,6,7,8,13,15,17,20,23,24,25,26,28), Hybrid (16,18,32,34,35,36) [corrected from an earlier version of this list, which had misclassified Hyundai NEXO (idx. 3), Mercedes T-Diesel (idx. 7), Mercedes GLC Coupe (idx. 32), and Renault Arcana (idx. 34), inconsistent with the fleet composition reported in Section 2.1].

After Agglomerative clustering in the first cluster comprised cars with indexes (3,5,7,8,10,12,13,14,16,21,23,25,26,27,28,33,34,35,36) in the second cluster (1,4,11,15,17,18,19,20,22,24,30,31,32,37) and in third cluster (2,6,9,29).

For DBSCAN in the first cluster we found cars with indexes (1,2,3,5,6,7,8,9,10,11,12,13,14,15,16, 20,21,23,25,26,27,28,29,30,31,33,34,35,36,37), for the second (17,18,22,24), and for the third (4,19,32).

For Gaussian clustering–1 cluster (1,3,4,5,10,11,14,15,18,19,20,22,23,24,25,26,28,33,37), second cluster (13,16,27,30,34,35,36) and third cluster (2,6,7,8,9,12,17,21,29,31,32).

For K-Means clustering in the first cluster we have (1,3,4,5,11,14,17,18,19,20,21,22,23,24,26,28,32,33), in the second (8,10,13,15,16,25,27,30,34,35,36,37) and in the third (2,6,7,9,12,29,31).

From just inspecting the clusters: K-Means and Gaussian Mixture seem most similar—both cluster (2,6,7,9,12,29,31) together as a cluster, and both frequently group (1,3,4,5,11,14,15,18,20,22,23, 24,25,26,28,33,37) together. Agglomerative and K-Means overlap somewhat, especially for clusters that include (1,3,5,14,23,26,33). DBSCAN is most different—its first cluster includes almost all points, acting more like a dense core grouping.

We estimated pairwise similarity between clustering results by comparing how many pairs of cars are clustered together in each pair of methods. To do that we calculated the Rand Index by counting how many pairs of elements (i, j) are in the same cluster in both A and B (SS) and in different clusters in both A and B (DD), then we computed:

$$Rand Index={\displaystyle \frac{SS+DD}{Total Pairs}}$$

(6)

A similarity score between 0 and 1 indicating how similarly the two algorithms clustered the data is presented in

Table 8. A similarity score between 0 and 1 indicating to what extent the data is clustered at 3 clusters clustering.

	ENGINETYPE	K-MEAN	GAUSSIAN	AGGLOMERATIVE	DBSCAN
ENGINE TYPE	1	0.532	0.566	0.505	0.465
K-MEANS	0.532	1	0.737	0.594	0.426
GAUSSIAN	0.566	0.737	1	0.560	0.419
AGGLOMERATIVE	0.505	0.594	0.560	1	0.552
DBSCAN	0.465	0.426	0.419	0.552	1

.

The results of the three-cluster solution obtained using K-Means, Gaussian Mixture, Agglomerative, and DBSCAN clustering are visualized in Figure 13.

The clusters overlap significantly, but each algorithm defines them differently due to their inherent logic. Most similar clustering algorithms (in order of similarity): K-Means & Gaussian Mixture—high overlap in cluster structure. Agglomerative & K-Means—moderate overlap. Agglomerative & Gaussian—some overlap. DBSCAN & others—least similar to all others; DBSCAN clusters are broader or more noise-tolerant.

Also, neither of the clustering method overlapped highly with the expected three clusters (ICE, Hybrid, EV). The highest similarity was obtained with the Gaussian Mixture Model (0.566), which is insufficient. From this we can conclude that the engine type factor is not of great importance for the current way in which the studied cars are naturally clustered differently, and that we need to look for another, currently hidden factor that influences more how the different cars are perceived by the driver.

Clustering into 8 clusters.

Based on the Theta/Alpha EEG increase or decrease, pulse rate increase or decrease and GSR increase or decrease we subdivided the vehicles in eight categories: CL1 (1,11,13,18,22,24), CL2 (5,6,12,14), CL3 (2,4,19,20,31,32), CL4 (9,29,33), CL5 (7,16,23,34), CL6 (3,10,21,26,27), CL7 (15,17,30,37), CL8 (8,25,28,35,36).

After Agglomerative clustering cars the following indexes fell into the relevant clusters CL1(15,17,30,37), CL2(8,13,25,35,36), CL3(7,16,23,34), CL4(3,10,21,26,27), CL5(2,4,19,20,31,32), CL6(9,29,33), CL7(1,11,18,22,24,28), CL8(5,6,12,14).

With DBSCAN we subdivided the car indexes as follows: CL1(1), CL2(2), CL3 (3,4,5,7,8,9,10,11,14,15,16,17,18,19,20,21,22,23,24,25,26,28,29,30,31,32,33,34,36,37), CL4(6), CL5(12), CL6(13), CL7(27), CL8(35).

After Gaussian clustering we obtained CL1(3,5,14,28,33), CL2(16,35,36), CL3(2), CL4(10,23,26,34,7), CL5(8,9,12,13,21,25,27,29,31), CL6(4,15,17,18,19,24,30,32,37), CL7(6), CL8(1,11,20,22).

For K-Means clustering our clusters were CL1(3,5,14,28,33), CL2(16,35,36), CL3(2), CL4(10,23,26,34), CL5(7,8,9,12,13,21,25,27,29), CL6(4,15,17,18,19,24,30,31,32,37), CL7(6), CL8(1,11,20,22). The resulting similarity scores are presented in

Table 9. A similarity score between 0 and 1 indicating to what extend the data is clustered at 8 clusters clustering.

	PHYS. STATES	K-MEAN	GAUSSIAN	AGGLOMERATIVE	DBSCAN
PHYS. STATES	1	0.809	0.800	0.812	0.375
K-MEANS	0.809	1	0.973	0.868	0.428
GAUSSIAN	0.800	0.973	1	0.841	0.419
AGGLOMERATIVE	0.812	0.868	0.841	1	0.455
DBSCAN	0.375	0.428	0.419	0.455	1

.

Unsupervised clustering into eight clusters showed that agglomerative clustering comes closest to the initial expectation of how the cars should be distributed. In addition, the table shows that the other two methods K-Means and Gaussian also give high similarities to the expectations and only DBSCAN subdivides the cars in a different way as presented on Figure 14.

We have also performed multidimensional scaling by using the resulting similarity matrix for all of the 37 cars tested. The visualization of the scaling is presented on Figure 15. We have found that the quality of the solution (Kruskal Stress-1) is that with two dimension Stress 1 index is 0.238 or poor—expected for 37 objects and with three dimension Stress 1 index is 0.135 or Fair/Good. With 37 objects, a value of around 0.24 in two dimensions is typical—the reduction to 0.14 in three dimensions the result confirms that the third dimension carries additional information. Working with the three dimension solution is recommended. The key observations are expanded on below.

In the two dimensional map (Dimension 1 × Dimension 2), several clear groupings emerge. On the right side of the lower quadrant, Hyundai NEXO and VW ID BUZZ are very close together—indicating a similar neurophysiological profile. In the upper right corner, Porsche Taycan, Opel Astra GSI, and BMW M2 form a cluster associated with high emotional activation. Strong outliers are DACIA JOGER and MERCEDES GLC, both positioned in the lower left corner—displaying a very different profile compared to the remaining objects. KIA EV9 and Renault Arcana are nearly overlapping, which is noteworthy given their very different market positioning—they likely elicit a similar autonomic response in respondents. In Figure 16 an analysis of the Positioning Map (Dimension II × Dimension III) is presented.

The map shows the projection of the MDS solution onto the second (20.9% explained variance) and third (16.6%) dimensions, together with attribute vectors. We can provide the following axes interpretation. Axis II (horizontal, 20.9%) is dominated by the PULSE vectors (PRE, DRIVE, POST), pointing to the left. This means that brands positioned on the left side elicit higher heart rate—greater physiological arousal. The axis can be interpreted as “physiological activation/stress”.

Axis III (vertical, 16.6%) is dominated by the skin-conductance (galvanic skin response) vectors pointing upward, and the EEG and SpO₂ (oxygen-saturation) vectors pointing downward. The upper zone corresponds to higher electrodermal activity, while the lower zone corresponds to higher EEG power and higher oxygen saturation. The axis can therefore be described as “electrodermal activity vs. cortical/oxygenation activity”. These are physiological descriptions only; no subjective or affective measures were collected. Regarding the quadrant analysis, we found the following groups.

Upper right quadrant is with high SC, low PULSE:

The SC PRE, SC DRIVE, and SC POST vectors point directly into this region. Sessions located here showed a strong electrodermal response without concurrent heart-rate elevation. This quadrant includes Hyundai NEXO, VW ID BUZZ, Hyundai STARIA, MUSTANG MACH1, DACIA JOGER, and SUBARU SOLTERRA. Without concurrent subjective measures, the affective meaning of this physiological pattern cannot be established.

Upper left quadrant is with high SC + high PULSE:

The MERCEDES GLC session was a strong outlier, combining high electrodermal response with elevated heart rate—a pattern of high overall autonomic activation. Without a concurrent subjective measure, this cannot be interpreted as reflecting ‘excitement’ versus ‘stress’; it indicates only that this single session elicited unusually high sympathetic activation on both channels.

Lower left quadrant is with high PULSE, low SC:

HAVAL DARGO, BMW M2, VW ARTEON, Alfa Giulietta, KIA EV9, Renault Arcana, Hyundai IONIQ 6, and MERCEDES AMG EQE SUV—these sessions showed elevated pulse with a weak skin-conductance response. In the absence of subjective ratings, this physiological pattern cannot be labelled as tension, anxiety, or excitement.

Lower right quadrant is with high EEG + SpO₂, low PULSE:

The VOLVO XC60 session was an outlier at the bottom of this projection, dominated by the EEG DRIVE, EEG POST, and SpO₂ DRIVE vectors, indicating comparatively high EEG power and oxygen saturation during and after driving with low cardiovascular loading. This is a physiological observation only and was not linked to any subjective measure.

Central zone quadrant (around the origin):

A large cluster of vehicles is concentrated near the center—Porsche Taycan, BMW iX, POLSTAR 2, Mercedes GLC Coupé, PEUGEOT e 2008, HONDA E NY1, VW Touareg, INEOS GRENADIER, MERCEDES S580, Smart Brabus, VW ID3, BMV XM, and others. These sessions display a mediocre, undifferentiated neurophysiological profile along axes II and III—they do not stand out meaningfully on either attribute vector.

The key findings from the analysis are the following:

The MERCEDES GLC session was the only one with simultaneously high activation on both axes.

The VOLVO XC60 session was an outlier on the EEG/oxygen-saturation dimension; this is a physiological observation only and was not linked to any subjective or preference measure.

Most sessions clustered near the centre of the map, indicating limited differentiation in their neurophysiological profile across these two dimensions.

Hyundai NEXO and VW ID BUZZ remain close to each other in this projection as well—a consistently similar profile across all three dimensions, which is a strong signal of competitive overlap.

4. Discussion

4.1. Vehicle Type Effects on Cortical Activity

The most robust and theoretically significant finding of the present study is the main effect of vehicle propulsion type on broadband EEG power spectral density, confirmed for three of the four frequency bands—theta, low beta, and high beta—across both the nonparametric and the repeated-measures-adjusted mixed-effects analyses. ICE vehicles were associated with significantly higher theta, low beta, and high beta power than EVs (all post hoc p ≤ 0.025); the ICE–Hybrid and EV–Hybrid contrasts did not reach significance, leaving the Hybrid group statistically intermediate. The alpha band showed the same pattern under the Kruskal–Wallis and post hoc tests (p_adj = 0.025), but this effect did not survive the repeated-measures-adjusted model (p = 0.107; Section 3.1.6); because that model is the more appropriate test for this single-driver, repeated-measures design, the alpha result should be treated as considerably less robust than the other three bands. This result—robust across complementary analyses for theta, low beta, and high beta—provides the first empirical evidence from simultaneous multimodal real-world recording that the propulsion environment of the vehicle cabin systematically modulates cortical oscillatory activity during driving.

The elevation of ICE broadband EEG power is consistent with—but not directly demonstrated to be caused by—the richer sensory environment typically associated with combustion cabins. Engine-generated low-frequency noise, drivetrain vibration, and exhaust harmonics are plausible contributors that could activate auditory and somatosensory cortical processing systems and engage the ascending reticular arousal system, producing elevated broadband cortical tone relative to the quieter EV environment [25,26]. However, acoustic, vibrational, and cabin air-quality variables were not directly measured in this study, and the tested vehicles also differed in mass, size, and cabin design beyond propulsion type; this explanation should therefore be regarded as one plausible mechanism among several confounded with vehicle group, rather than an established cause (see Section 4.7). The gradient of significance across bands—with the largest H statistics in the beta range (β High: H = 18.33)—is consistent with the established role of beta oscillations in sensory arousal, active processing, and emotional engagement [38]: the acoustically richer ICE cabin produces greater sensorimotor cortical activation precisely in the spectral range most sensitive to external stimulation.

The absence of any significant EV–Hybrid difference, despite the operational differences between these vehicle types, is ecologically meaningful. Modern hybrid vehicles operate primarily on electric power at the speeds characteristic of the standardized urban-suburban driving route employed in this study, producing a sensory environment acoustically similar to full EV operation. This acoustic convergence may explain the functional equivalence of EV and Hybrid EEG profiles and has practical implications for NVH engineering: the neurophysiological benefits of a quieter cabin may be substantially realized in current hybrid technology without full electrification.

4.2. ICE-Specific Alpha Decline and Anticipatory Arousal

A theoretically important qualitative Group × Phase interaction was observed in the alpha and beta bands: ICE sessions showed the highest PREDRIVE-phase power with a monotonic decline across DRIVE and POSTDRIVE (ICE alpha: PREDRIVE 2.07 → DRIVE 1.17 → POSTDRIVE 0.95 µV²/Hz), while EV and Hybrid sessions showed comparatively flat or U-shaped phase profiles. This ICE PREDRIVE-phase alpha elevation—substantially above that of EV and Hybrid groups—may reflect pre-existing elevated arousal or anticipatory autonomic activation in the period immediately preceding the drive. For experienced ICE drivers, the pre-drive baseline may itself be characterized by a qualitatively different cognitive and affective state than for EV or Hybrid drivers, potentially including higher task readiness, sensory anticipation, or driving-related anxiety. The progressive alpha decline during and after driving would then represent habituation of this elevated baseline state rather than task-induced suppression per se. This interpretation parallels the anticipatory arousal explanation proposed for the GSR findings (Section 4.3) and underscores the importance of characterizing pre-drive baselines as dynamic psychophysiological states rather than neutral reference conditions.

4.3. Phase Effects: Theta as a Workload Biomarker

The selective theta-band elevation during active driving—significant at the omnibus level and confirmed within both ICE and EV groups by Friedman tests—is consistent with a substantial body of literature identifying frontal midline theta as the most reliable spectral index of cognitive workload during driving [33,34,37]. The theta response was phasic and task-specific: it did not persist into the POSTDRIVE resting state (PREDRIVE vs. POSTDRIVE: W = 146, p = 0.468), arguing against slow state drift or circadian confounds and supporting the interpretation that theta augmentation reflects genuine task-coupled cognitive engagement. An additional, more marginal EV-specific phase modulation in high beta (χ²(2) = 6.00, p = 0.050, not surviving in low beta once NEXO is correctly classified as EV) is consistent with a mobilization-and-release of cortical arousal across the session, potentially reflecting the attentional novelty of the perceptually quieter EV environment or the driver’s active monitoring of unfamiliar vehicle dynamics—though this interpretation should now be treated as tentative given the weaker statistical support.

4.4. Autonomic Findings: The Paradox of Driving-Related HRV Elevation

The most counterintuitive finding of the present study is the robust DRIVE-phase elevation in HRV-RMSSD across all vehicle groups (~34% above PREDRIVE baseline, Kruskal–Wallis H(2) = 8.66, p = 0.013; Friedman tests significant in both EV and ICE groups independently). Increased RMSSD during active driving contradicts a simple stress-arousal model, under which sympathetic activation and cognitive demand would be expected to suppress vagal modulation. Three non-mutually exclusive interpretations are offered.

First, the elevated RMSSD during DRIVE may reflect task-induced respiratory regulation. Active cognitive engagement during driving may produce a characteristic pattern of slowed, deeper respiration associated with focused attention—a well-documented respiratory adjustment during sustained cognitive effort [39]—which would mechanically increase respiratory sinus arrhythmia and thus RMSSD through cardiorespiratory coupling, independent of any change in autonomic outflow. Second, the pre-drive PREDRIVE baseline may have been characterized by anticipatory sympathetic arousal that suppressed vagal tone below the driver’s true resting level, making the DRIVE phase appear comparatively elevated. This interpretation is strongly supported by the parallel GSR finding—where PREDRIVE was the highest arousal phase across all sessions—and aligns with the alpha elevation observed at PREDRIVE in the ICE group. Third, for an experienced professional driver, the driving task itself may function as a form of engaging but relaxing occupation, reducing ruminative cognitive activity relative to the aroused waiting state preceding the drive.

The HR–SDNN correlation was strongly negative at PREDRIVE (rs = −0.837) but substantially attenuated during driving (DRIVE: rs = −0.488), indicating that the robust sympathovagal relationship characterising resting physiology was disrupted by task-induced autonomic dynamics during driving. The DRIVE-phase emergence of a significant negative HR–RMSSD correlation (rs = −0.415) further supports the interpretation that active driving specifically engaged a vagal-sympathetic competitive dynamic not present at rest.

A fourth possibility is that the elevated RMSSD reflects a mild vestibular/visual mismatch response during active driving rather than pure workload or anticipatory-arousal resolution. Molefi et al. [40] reported HRV changes associated with visually induced motion sickness, raising the possibility that active driving—involving continuous visuo-vestibular integration absent during seated baseline—modulates vagal tone through mechanisms related to, but distinct from, cognitive workload. This interpretation would predict individual differences in RMSSD response tied to motion-sickness susceptibility, which future multi-participant studies could test directly.

4.5. GSR: A Universal Anticipatory Arousal Signature

The GSR findings are remarkable for their consistency and magnitude. A strong, highly significant Friedman phase effect was found in all three vehicle groups (EV χ²(2) = 17.38, ICE χ²(2) = 24.40, Hybrid χ²(2) = 12.00, all p ≤ 0.003)—meaning that PREDRIVE > DRIVE > POSTDRIVE electrodermal activity held in nearly every session—and the whole-sample PREDRIVE–POSTDRIVE reduction (W = 8.0, p < 0.001, ~37% decline) was among the strongest effects in the study. This universality argues strongly against vehicle-type as a determinant of electrodermal arousal session trajectory: the pre-drive baseline was consistently the highest arousal state regardless of which vehicle type (EV, ICE, or Hybrid) was subsequently driven.

The most parsimonious account is that the pre-drive PREDRIVE period was contaminated by anticipatory sympathetic arousal—associated with equipment fitting, being monitored, and preparing for an experimental drive—that progressively habituated during and after driving. This interpretation reframes the session trajectory from a driving-induced stress response to a protocol-dependent arousal-habituation sequence. The absence of any group difference in GSR is consistent with this framing: anticipatory arousal is a function of experimental context rather than vehicle type.

4.6. SpO₂: No Vehicle-Type-Dependent Modulation

SpO₂ remained stable across the session. Recording phase had no significant omnibus effect (H(2) = 0.60, p = 0.741) and no significant PREDRIVE–POSTDRIVE change was detected. Within-group Friedman tests showed no significant phase effect in any vehicle group (EV p = 0.269, ICE p = 0.138, Hybrid p = 0.247), and the rank-based mixed-effects model found no significant Group × Phase interaction (χ²(4) = 6.43, p = 0.169; Section 3.2.4). Apparent ICE and Hybrid SpO₂ phase effects reported in an earlier draft were artefacts of the pre-correction vehicle classification and do not survive re-analysis. Peripheral oxygenation therefore showed no vehicle-type-dependent modulation; any residual between-phase differences were clinically negligible (well under 0.1 percentage points) and statistically non-significant. This null result should be interpreted with caution given the small, unbalanced sample and the limited dynamic range of fingertip pulse oximetry.

4.7. Limitations and Future Directions

Several methodological limitations must be acknowledged. Most critically, the study employed a single professional test driver rather than a heterogeneous participant sample, providing excellent within-driver experimental control but precluding generalization to the broader population of drivers. Because sessions were conducted on different days, intra-individual variability due to fatigue, sleep quality, circadian timing, and day-to-day mood could not be controlled and may covary with, rather than be independent of, vehicle type if session scheduling was not randomized with respect to vehicle group. The observed physiological differences are attributable to vehicle characteristics as experienced by this specific driver under this protocol, and future studies must employ larger, demographically diverse samples with crossover designs. The Hybrid group was small (n = 6 enrolled vehicles) and had insufficient complete-case triads for within-group Friedman analysis, substantially limiting interpretation of Hybrid-specific findings; this small, unbalanced group also reduces statistical power to detect true differences between Hybrid and the other groups and increases the risk of Type II error for any Hybrid-specific null result reported here (e.g., the non-significant EV–Hybrid EEG contrasts, Section 3.1.3)—non-significant Hybrid comparisons should not be interpreted as evidence of equivalence. Missing data in the HR/HRV dataset (13–31% per cell for EV/ICE; 33–50% for Hybrid) further reduced statistical power for cardiovascular comparisons.

A related limitation is that the tested vehicles were not matched or controlled for mass, size, power output, tyre specification, cabin design, or suspension type; these characteristics vary substantially within as well as between the EV, ICE, and Hybrid groups and are confounded with propulsion type in this convenience sample of 37 vehicles. The reported group differences should therefore be interpreted as associations between propulsion category (and its typical correlated vehicle characteristics) and driver physiology, not as evidence that propulsion type per se is the causal mechanism. A formal analysis adjusting for these covariates (e.g., curb weight, power output, tyre specification) was not conducted in this study; compiling such a dataset and testing whether the reported group differences persist after adjustment is identified as a priority for future replication work with a larger, better-characterized vehicle sample.

Methodologically, the use of a consumer-grade wireless EEG headset (EMOTIV EPOC X)—while appropriate for naturalistic on-road recording—provides lower spatial resolution and less artefact control than research-grade amplifiers with independent component analysis. Future studies should employ ICA-based artefact rejection, source localization, and formal tests for vibration-induced EEG contamination across vehicle types. The GSR anticipatory arousal confound highlights the need for extended pre-drive adaptation periods and repeated-session designs to characterize true resting baselines. Event-related physiological analyses tied to specific driving events (gear changes, braking, and intersections) would provide mechanistically richer characterization of within-drive dynamics.

Future research should prioritize: (1) larger, balanced, multi-participant crossover designs with standardized routes and extended baseline periods; (2) ICA-based EEG preprocessing and source analysis; (3) event-related physiological analyses; (4) concurrent subjective measures of workload, stress, vehicle preference, and EV-specific psychological factors such as range anxiety [41] (NASA-TLX, PANAS); (5) in-cabin air quality monitoring (CO₂, particulates) to test cabin environment hypotheses; and (6) longitudinal designs examining whether physiological adaptation to EV or ICE operation occurs across repeated sessions.

4.8. Clustering Findings

This study aimed to explore the underlying structure of vehicle-related data using unsupervised clustering techniques, with two primary hypotheses guiding the analysis: first, that the data would naturally cluster into three groups corresponding to vehicle powertrain types (ICE, Hybrid, EV); and second, that it may instead be better explained by eight psychophysiological states derived from post-driving biometric data.

The clustering analyses revealed that K-Means and Gaussian Mixture Models (GMM) produced the most similar cluster structures, with consistent groupings. Agglomerative clustering showed moderate overlap with both K-Means and GMM, while DBSCAN diverged significantly, yielding broad, noise-tolerant clusters with minimal alignment to the others.

Critically, none of the clustering methods aligned strongly with the a priori classification into ICE, Hybrid, and EV categories. The highest similarity, achieved by the Gaussian Mixture Model (RI = 0.566), falls short of a meaningful match, suggesting that engine type is not the principal factor shaping the natural structure of the data as experienced or perceived by drivers.

In contrast, clustering into eight groups, motivated by a theoretical psychophysiological model, yielded stronger support for latent structure in the data. Agglomerative clustering most closely matched the hypothesized eight-state framework, with K-Means and GMM also showing reasonable alignment. This suggests that drivers’ post-driving physiological responses—specifically changes in heart rate, EEG theta/alpha ratio, and skin conductance may underlie a more meaningful and consistent basis for differentiation among vehicle experiences than powertrain type alone.

These findings highlight the importance of considering human-centered physiological indicators in vehicle classification and evaluation. Future research should further investigate these hidden factors and their relationship to vehicle design, driver perception, and cognitive-affective responses during and after driving, an approach with precedent in neuromarketing research applying EEG and other physiological measures to consumer and brand perception [42,43].

4.9. Applied Implications

Despite the noted limitations, the findings carry implications for automotive design, human factors, and policy. The demonstration that ICE vehicle operation is associated with significantly elevated broadband cortical activity—a pattern consistent with greater sensory and cognitive load—suggests that vehicle propulsion type, through mechanisms that plausibly include but were not directly measured here (acoustic, vibrotactile, and air-quality differences), constitutes a meaningful correlate of driver neurophysiological state. If sustained across longer driving sessions and replicated in larger samples, higher cortical arousal in ICE drivers could contribute to faster-onset fatigue and attentional deterioration, with implications for long-haul road safety.

From an automotive engineering perspective, the equivalence of EV and Hybrid EEG profiles suggests that the neurophysiological benefits of a quieter cabin may be achievable in hybrid technology without full electrification, providing a near-term pathway to psychophysiological improvement across the transitional vehicle fleet. The present data also suggest that EEG-based driver monitoring systems—increasingly deployed in production vehicles—will require vehicle-type-specific calibration, since ICE and EV drivers exhibit substantially different baseline spectral profiles even in the pre-drive resting state. A universal theta threshold for workload detection would perform differently across vehicle types given the tonic spectral elevation in ICE drivers documented here.

5. Conclusions

This study provides the first simultaneous multimodal neurophysiological characterization of EV, ICE, and Hybrid vehicle driving under real-world conditions. The principal conclusions are:

• Vehicle propulsion type modulates broadband EEG power, robustly across three of the four spectral bands. A significant omnibus group effect was found in all four bands under standard nonparametric tests; ICE driving showed significantly higher theta, low beta, and high beta EEG power than EV driving under both the nonparametric and repeated-measures-adjusted analyses (all post hoc p ≤ 0.025), with the Hybrid group statistically intermediate and not significantly different from either. The alpha-band group difference (p_adj = 0.025) did not survive the repeated-measures-adjusted mixed-effects model (p = 0.107) and should be treated as the least robust of the four. This represents the study’s strongest finding for theta, low beta, and high beta.
• Theta oscillations index task-coupled cognitive workload during driving. A selective, phasic theta augmentation during DRIVE—significant across the full sample and confirmed within both ICE and EV groups—is consistent with established models of frontal-midline theta as a continuous neurometric of driver cognitive load.
• GSR declines universally across the session regardless of vehicle type. A strong, highly significant Friedman phase effect for GSR in every vehicle group (all p ≤ 0.003) indicates that PREDRIVE > DRIVE > POSTDRIVE was the rank order in nearly every session. This reflects anticipatory autonomic arousal at pre-drive baseline rather than stress accumulation during driving.
• HRV-RMSSD paradoxically increases during driving. A significant ~34% DRIVE-phase elevation in vagal cardiac modulation was observed across vehicle groups, most plausibly attributable to task-coupled respiratory regulation or resolution of anticipatory autonomic suppression.
• SpO₂ showed no significant group effect, phase effect, or Group × Phase interaction. Peripheral oxygenation was stable across vehicle types and session phases; the ICE-specific dip reported in earlier drafts did not survive correction of the vehicle classification.

The exploratory multidimensional-scaling analysis (Section 3.4) identified individual sessions (e.g., MERCEDES GLC, VOLVO XC60) with distinctive physiological profiles and a large central cluster of undifferentiated sessions, with Hyundai NEXO and VW ID BUZZ consistently close together. Because no subjective or preference measures were collected, these patterns are exploratory and hypothesis-generating only, and are not interpreted here in terms of brand image, pleasure, or emotional valence.

Collectively, these findings demonstrate that vehicle propulsion technology constitutes a significant and physiologically consequential determinant of driver cortical and autonomic state, with ICE operation imposing a higher cortical load than EV operation and the Hybrid group occupying an intermediate position. Replication with larger, diverse samples and crossover designs will be essential to establish the robustness and generalizability of these effects across the population of drivers currently navigating the global transition to electric mobility.