Driver Behavior Mechanisms and Conflict Risk Patterns in Tunnel-Interchange Connecting Sections: A Comprehensive Investigation Based on the Behavioral Adaptation Theory

Abstract

Tunnel-interchange sections are characterized by complex driving tasks and frequent traffic conflicts, posing substantial challenges to overall safety and efficiency. Enhancing safety in these areas is crucial for the sustainability of traffic systems. This study applies behavior adaptation theory as an integrated framework to examine the impact of environmental stimuli on driving behavior and conflict risk in small-spaced sections. Through driving simulation, 19 observation indicators are collected, covering eye-tracking, heart rate, subjective workload, driving performance, and conflict risk. The analysis, using single-factor ranking (Shapley Additive Explanation), interaction effects (dependence plots), and multi-factor analysis (Structural Equation Modeling), demonstrates that driving workload and performance dominate the fully mediating effects between external factors and conflict risk. High-load environmental stimuli, such as narrow spacing (≤500 m) and overloaded signage information (>6 units), significantly elevate drivers’ stress responses and impair visual acuity, thereby increasing task difficulty and conflict risk. Critical factors like saccade size, heart rate variability, lane deviation, and headway distance emerge as vital indicators for monitoring and supporting driving decisions. These findings provide valuable insights for the operational management of small-spacing sections and enhance the understanding of driving safety in these areas from a human factor perspective.

1. Introduction

In China, the rapid development of highways in mountainous regions has led to the construction of numerous tunnel and interchange sections with close spacing. These areas are marked by significant traffic conflicts and high crash rates, presenting challenges not only to road safety and efficiency but also to the overall sustainability of the highway system. The intricate and confined design of these sections significantly challenges drivers’ cognitive and response abilities due to frequent environmental changes and complex traffic merges [1,2]. These demanding conditions necessitate swift decision-making, often leading to critical driving errors such as missed exits, emergency braking, and sudden lane changes, thereby increasing the risk of crashes [3,4].

Given these complexities, it is essential to explore how external factors and internal workload influence drivers’ behavior and risk patterns in small-spacing sections, particularly from a behavioral perspective. Behavioral Adaptation Theory (BAT) offers a useful framework in this context, emphasizing the interaction between drivers’ mental workload and behavioral feedback under complex stimuli [5]. BAT suggests that driver behavior is shaped by a series of cognitive strategies and control mechanisms, not merely by simple rules [6,7]. This insight is essential for enhancing traffic safety in challenging environments like tunnels or interchanges, where demands on drivers are particularly high.

Despite extensive research on driving behavior within tunnel environments or interchange sections individually, the combined impact of these factors on driver behavior is less understood. Studies have demonstrated that abrupt changes in lighting at tunnel exits can severely impair drivers’ perception of time headway and velocity [8,9]. Additionally, limited sight distance and visual biases can lead to increased heart rate fluctuations, enlarged pupil diameters, and lane deviations, which result in higher mental loads and degraded performance [10,11,12]. In interchange diverging sections, drivers must navigate complex maneuvers like deceleration, turning, and lane changing, often while contending with distracting external stimuli [13,14,15]. Furthermore, the necessity for drivers to transition from the inner lane to the outer lane results in frequent and concentrated mandatory lane changes, significantly contributing to traffic conflicts [4].

However, the unique and combined effect of small spacing between tunnels and interchange exits presents additional, more complex challenges for drivers. Although previous studies have provided valuable insights into the effects of visual differences and diversion environments, they may not fully capture the nuanced behavioral dynamics of tunnel-interchange connection sections with limited spacing. Given the restrictions on lane changing within tunnels, off-ramp vehicles must quickly adapt to environmental changes, interpret signage, and find acceptable gaps in limited distances. Spatial limitations and high task demands in these areas make driver decision-making more dynamic and challenging [9]. This indicates that conclusions drawn from broader contexts may not directly apply to small-spacing sections.

Addressing the complexity of these behavioral mechanisms—encompassing factors like information stimuli, subjective states, and driving performance—requires multifactorial analysis methods. Among this, Structural Equation Modeling (SEM) [16,17,18] and SHapley Additive explanations (SHAP) [19,20] are noteworthy, offering complementary methodologies for understanding driving behaviors in complex scenarios. SEM serves as a valuable parametric method widely used to estimate latent variables, enabling an in-depth exploration of mediating relationships and indirect effects between various driving factors [21,22,23]. Conversely, SHAP complements this by quantifying how individual features influence outcomes, directly attributing effects to observed indicators [24]. Notably, this method accounts for the interdependence among variables, which has been particularly useful in investigating significant factors affecting driving fatigue and performance [25,26]. The combination of these two methods allows for a comprehensive understanding of behavioral mechanisms in complex scenarios, enhancing our ability to uncover latent behavioral risks and adaptive strategies.

This study aims to bridge gaps in existing literature by employing a comprehensive modeling approach that integrates cognitive workload, environmental factors, and driver performance, ultimately enhancing our understanding of how drivers adapt to complex and dynamic roadway conditions. Specifically, the study will collect vehicle control and driver physiological data through simulated small-spacing scenarios to investigate how environmental factors influence driving characteristics in terms of workload, behavioral performance, and conflict risk. The findings will offer a scientific basis for the planning and design of tunnel-interchange sections with small spacing, thus enhancing road safety.

2. Theoretical Background

BAT can be viewed as a set of concepts that highlight the intricate relationships between environmental and driver-related factors, directly involving driving performance and overall safety [27]. To probe the underlying mechanism, this study incorporates three fundamental concepts in BAT drawn from information processing, cognitive psychology, and risk adaptation: Stimulus-Response (S-R) mechanism, Driving Workload Theory (DWT), and Risk Allostasis Theory (RAT) [14,28,29,30]. Figure 1 illustrates this integrated paradigm model. This section presents the review of these theories and proposes the theoretical assumption for this study.

2.1. S-R Mechanism

The S-R mechanism, a well-established model in behavioral psychology, explains how adaptive individuals respond to external stimuli within complex systems [31], positing that behavioral adaptation aims to limit or avoid risk [32]. This model suggests that an individual’s behavior results from the interplay between stimulus and response, with the response comprising specific physiological feedback (visual, cognitive, and psychological) that informs behavioral performance. For advanced driving tasks, intense external stimuli can disrupt drivers’ information processing, leading to degraded performance [33]. While earlier studies depicted driver behavior as a sequence of independent decisions driven by expectations [32], critiques have emerged for failing to capture complex behaviors like overtaking or merging [34].

To explain this limitation, Toates identified interactions involving S-R and cognitive processes that jointly regulate behavior [35]. Subsequent models have refined this mechanism to Stimulus- Cognition- Response feedback, enhancing our understanding of how driving behavior adapts to the external environment in traffic systems [36,37]. Recent studies have further enhanced this model by linking subjective cognition with feedback mechanisms and applying information theory and cognitive psychology to analyze how traffic information densities impact driving errors and visual memory load [14]. Such insights are particularly relevant in dynamic driving environments like tunnels and interchanges, where frequent environmental changes demand high levels of cognitive and visual adaptation [12,13,38]. These findings underscore the role of cognitive processes in shaping driver behavior in complex scenarios.

2.2. Mechanism and Measurement of Driving Workload

Driving workload, assessed as a measure of driving demands and task stress, refers to the effort required to keep driving within a subjective safety interval. The Yerkes–Dodson law suggests that an optimal workload level exists for each task, where both excessive and insufficient workloads can impair performance [39]. The hierarchical “Demand-Workload-Effort” theory, or DWT, explains the relationship between subjective demand, workload, and performance [40]. According to DWT, as subjective demand increases, performance can also improve—up to a certain point—after which excessive demand overwhelms the driver and negates any benefits. Complex or monotonous driving scenarios can cause mental overload or underload, potentially leading to dangerous driving behaviors or distraction, consistent with information processing and attention restoration theories [41,42].

The compatibility of DWT makes it a foundational framework for interpreting the human factors involved in driving. It is linked to mental constraints and physiological responses, typically measured through both subjective and objective indicators. Subjective measures include tools like the NASA Task Load Index (NASA-TLX) and the Subjective Workload Assessment Technique to standardize mental, temporal, and task demands [17]. In contrast, objective measures involve physiological indicators such as eye movements, electrocardiography (ECG), brain waves, and muscle responses [30]. Since about 80% of driving information is visual [43], metrics such as fixation, saccade, and pupil diameter are crucial for assessing cognitive workload. Additionally, heart rate variations (HRV) serve as reliable indicators of task difficulty [21]. These workload measurements help quantify the psychological cost of specific tasks, reflecting the complexity of driving challenges.

2.3. RAT

Proposed by Fuller in 2005, RAT evolves from the motivational risk model and broadly applies to risk-based decision-making [44]. It posits that individuals strive to align their perceived risk with their internal expectations as closely as possible, adjusting their behavior to maintain dynamic equilibrium when significant deviations occur. This adjustment process is described by the Task Capacity Interface (TCI) mechanism, which merges competition theory with RAT, emphasizing the interplay between perceived risk and task demand (Figure 1) [45]. Under this framework, perceived risk serves as an indicator of task difficulty and a primary determinant of driver behavior [46].

The fundamental premise of the motivational model posits that individuals possess a preferred range of risk perceptions and adjust their behavior accordingly to manage task difficulties [47]. If cognitive demand exceeds their capacity, they may struggle to respond properly to task requests, leading to potential conflicts or crashes. Under the RAT framework, the range of risk acceptance is more flexible with mental workload and external environment. This flexibility is particularly relevant to the present study, as it relates subjective states to objective risks, accounting for both driving performance and errors.

Recent empirical evidence supports a dynamic relationship between risk perception, task difficulty, and driving performance, showing that increases in driving metrics correlate with greater task difficulty and risk perception, while the risk of loss of control follows a threshold pattern [48,49]. Once a driver exceeds their preferred speed, risk perception and task difficulty tightly correlate with statistical risk assessments [46]. This mechanism underpins various behavior and risk control strategies, including car-following models and adaptive cruise control [50,51]. Therefore, to enhance the applicability of RAT, this study integrates the concepts of perceived risk and task demand into the DWT framework, using observable conflict indicators instead of relying solely on subjective error rates.

2.4. Theoretical Framework

S-R, DWT, and RAT provide complementary perspectives for interpreting driving behavior in complex traffic systems. Firstly, these models elucidate behavioral adaptation by emphasizing factors such as stimulus intensity, mental load, and cognitive demand that shape the behavioral response mechanism. Secondly, the concept of “stimulus-cognition” highlights an information transfer process, enhancing our understanding of how the environment indirectly influences risk perception and driving errors. Furthermore, driving workload serves as a moderator of subjective factors within both the S-R and RAT models, enabling these theories to form a cohesive loop.

Given the above analysis, the underlying hypothesis of this study posits that the external environment indirectly influences driving performance and conflict risk by altering driving workload through external stimulation. Specifically, the hypotheses are as follows: (1) external stimuli directly affect both subjective and objective driver workload; (2) driver workload directly influences driving performance, including both lateral and longitudinal aspects; (3) driving performance directly impacts conflict risk; and (4) the influence of external stimuli on conflict risk is mediated by the relationships described in hypotheses (1) to (3).

To explore this, a theoretical framework with six latent variables has been developed, as illustrated in Figure 2. Notably, this framework avoids a strict hypothesis form to facilitate the exploration of multifactorial associations without being hindered by the complexity of numerous observed variables. Therefore, this study applies SHAP analysis with machine learning models to validate the theoretical model by feature importance and partial dependence. After the preliminary analysis, the PLS-SEM model was used to investigate the direct and mediating effects between the composite constructs in Figure 2.

The proposed framework and methodologies extend the application of BAT to the context of tunnel-interchange sections, offering new insights into the adaptive mechanisms drivers employ in these environments. This integration of S-R, DWT, and RAT contributes to advancing theoretical understanding while providing practical implications for improving road safety in challenging traffic environments.

3. Methods

To investigate the driving performance and conflict risk under the complex scenarios, the methodological steps are described through a simplified flow chart in Figure 3, including three parts: (1) driving simulation design; (2) data processing (3) statistical modeling. All the methodological steps are explained in detail in the following subsections.

3.1. Driving Simulation Design

Adequate and valid experiment data are a necessary prerequisite for driving performance assessment. However, limitations with engineering samples often hinder the verification of impacts in tunnel-interchange scenarios through field tests. Recent studies have shown that driving simulations are effective tools for human factor analysis [10,21]. To ensure safety and reproducibility, this study uses simulations to evaluate the physiological state and driving performance of drivers navigating tunnel-interchange segments.

3.1.1. Scenarios Design

To accurately recreate the driving scenarios, an extensive survey of 11 road samples in China was conducted. The UC-win/road 13.0 platform was employed to create the simulation environments. Each test section was divided into four parts: adaptation, tunnel, connection, and divergence. The 500 m area preceding the diversion point was designated as the data collection zone to capture drivers’ early lane-changing behaviors. The road layout and facility arrangements are detailed in Figure 4.

Based on the S-R framework for exogenous stimuli, four parameters were selected as independent variables:

(1) Connection distance (CD): Set for 100 m, 300 m, 500 m, 700 m, and 1000 m. Studies have shown that small clearances affect drivers’ reaction time and may lead to higher failure rates [3,4]. Therefore, CD is considered a potential factor contributing to workload.

(2) Tunnel length (TL): Configured at 500 m, 1500 m, and 3000 m to represent medium, long, and extra-long tunnels, respectively. Longer tunnels have been shown to increase cognitive and concave loads, underscoring the need for controlling tunnel dimensions.

(3) Information volume (IV) of traffic signs: Traffic sign layouts were based on engineering examples from the G5 Jing-Kun and G30 Lian-Huo expressways in Shaanxi Province. Following methodologies tailored to the Chinese context, blocks or symbols were counted as individual information units. The division of information units and the arrangement of signs are illustrated in Figure 5.

(4) Traffic volume (TV): Recognized as a fundamental variable linked to traffic conflicts, realistic traffic flow is vital for accurate simulations. Typical traffic flows in tunnel-interchange sections range from 360 to 1240 pcu/h. For this study, traffic volumes of 500, 800, and 1100 pcu/h were chosen, corresponding to service levels I to II.

The design elements of the roads are consistent across all simulation scenarios, with the essential details summarized in

Table 1. Control variables selection and basic roadway design information.

(a) Control variables selection
Variable	Value	Variable	Value
CD (m)	100, 300, 500, 700, 1000	IV (unit)	6, 9, 12, 15
TL (m)	1000, 1500, 3000	TV (pcu/h)	700, 900, 1100
(b) Roadway design information
Design speed (km/h)	100	Limit speed (km/h)	Mainline: 100, ramps: 60
Road base design	Integral road base	Number of lanes	Mainline: 3, ramps: 1
Shoulder width (m)	3.0	Lane width (m)	3.75
Median width (m)	3.0	Transition length (m)	100
Deceleration lane (m)	150	Transition ratio	1/40

. A total of 60 (3 × 4 × 5) road configurations under three traffic scenarios are prepared for simulation, with traffic volume settings adjusted before each test.

3.1.2. Simulation System

The simulated system and scenarios are shown in Figure 6. The driving simulation system consists of four parts: 6-degree-of-freedom (DOF) motion simulator, UC-win/road 13.0 simulation platform, Biopac MP160 electroconductive physiological recorder and SMI eye-tracking device. The motion simulator can simulate six DoF to recreate realistic driving conditions, such as acceleration, deceleration, steering, and sideslipping, and replicate vehicle vibrations and road bumps. The visual system of the simulator is equipped by three HD screens with 130° horizontal view and 40° vertical view. Vehicle operation data are collected at 100 Hz into the UC-win/road 13.0 and is integrated with the operation of surrounding vehicles for calculating driving performance and conflict metrics.

Real-time electrocardiogram (ECG) and eye movement data are collected from the driver and imported into the simulation database. The MP160 unit is used to collect ECG at 1000 Hz, and heart rate indicators are exported via AcqKnowledge 5.0 software. Eye movement parameters, including blink, saccade, fixation, and pupil diameter, are collected at 250 Hz by SMI (SensoMotoric Instruments, Teltow, Germany) oculography and combined with Begaze 3.7 for quantitative analysis.

The NASA-TLX is used to calibrate the subjective workload of drivers in this study. The results are collected and summarized in a questionnaire after each trial.

3.1.3. Samples and Participants

In this study, each driver’s diverging process is treated as an independent sample for modeling evaluation. To ensure robust analysis, a significant number of samples is needed, directly linked to participant count. The sample size was determined based on multiple factors. According to SEM sample size guidelines by Hair et al., models with more than 6 factors require at least 500 sample data points [23]. For machine learning (ML) models, larger datasets typically yield more accurate results [52]. Considering that our final model involved numerous factors, with 25 drivers completing 60 diverging scenarios each, we can obtain a total of 1500 valid diverging samples, which is sufficient for statistical power and reliability in both SEM and machine learning models. Additionally, practical constraints, such as resource availability and participant recruitment capacity, were considered when determining the sample size.

A total of 25 drivers (14 males and 11 females) were recruited to maximize the sample size while addressing logistical constraints. Participants were recruited through local advertisements and personal contacts, and participation was entirely voluntary. Each participant provided written informed consent prior to participation. The study was reviewed and approved by the Ethics Committee of the Highway School at Chang’an University, ensuring compliance with all relevant guidelines. Participants were informed of the general nature of the test purpose and driving simulation, but specific research hypotheses were not disclosed to prevent potential bias in their behavior.

To control for individual differences, participants were selected for their similar age, education, and driving experience. All the participants are aged 26~33 years (Mean = 29.38, SD = 1.78), with 3~7 years of driving experience (Mean = 3.97, SD = 1.01). Additional criteria include an average annual driving mileage of at least 20,000 km, a valid driver’s license, no more than one traffic violation in the past year, visual acuity greater than 5.0 (as per the new National Standard Vision Scale in China), and no physical impairments or mental health conditions that could affect cognitive abilities. None of the participants had prior experience with similar simulation experiments.

3.1.4. Experiment Procedure

Similarly to our previous study, the simulation experiment comprises three phases: preparation, pre-test, and test [3,53]. In the preparation phase, staff fine-tune the equipment, upload experimental scenarios, and set traffic volume configurations. During the pre-test phase, participants enter the simulated cabin, equip physiological acquisition instruments, and establish baseline measurements of heart rate and pupil diameter in a calm state. Before each test, participants are briefed on the driving task and destination and undergo a pre-drive to familiarize themselves with the simulator’s controls. To ensure focus during sign recognition, the ratio of diverging to driving straight is maintained at 3:1. In addition, drivers are instructed to adhere strictly to traffic signs, markings, and regulations, avoiding any actions unrelated to driving.

In the test phase, drivers operate the simulator without external interference. Immediately following each driving task, participants complete the NASA-TLX to minimize errors in subjective demand assessments. After a short relaxation break, the next experimental scenario is initiated.

The above process was rotated in each experimental scene, and after completing five scenarios (approximately 20–30 min), the driver entered the rest room while the next participant was tested. The driving simulations were conducted from 8:00 to 11:30 and 2:00 to 17:30 each day, with three participants in turn through each time period. The process balances the efficiency of the experiment with the necessary rest time for the participant. Importantly, no participants reported experiencing symptoms of simulator sickness, likely due to the relatively short duration of exposure and the provision of adequate rest intervals.

3.2. Data Processing

The driving operation and physiological data among the diverge section are collected for each experiment. The following section describes the processing and aggregation of raw data with several appropriate indicators to represent driving workload, performance, and conflict risk.

3.2.1. Data Preprocessing

To ensure the reliability and quality of the dataset, the data preprocessing involved the following steps:

(1). Invalid data removal: Collected datasets with significant noise, sensor errors, or missing critical values were eliminated. Criteria for invalid data included missing values exceeding 5% of a record, non-physiological values (e.g., negative pupil diameters), or incomplete trip records. After removing the invalid data, a total of 1482 sets are collected.

(2). Outlier detection: Extreme outliers were identified using the interquartile range (IQR) method, with values exceeding 2 times the IQR from the first or third quartile considered as outliers. This procedure was applied to continuous variables (e.g., vehicle speed, heart rate, fixation duration) for each driving test. Outliers were either corrected (if they were attributed to identifiable measurement errors) or removed if no rational correction could be made.

(3). Missing data: For data records with isolated missing values (<5% of the data), linear interpolation was used to estimate the missing values, ensuring minimal disruption to the data integrity.

(4). Data standardization: To facilitate comparisons and ensure model convergence in SEM, data standardization was performed using Z-score normalization, which ensured that all indicators had a mean of 0 and a standard deviation of 1. Notably, this step was applied only for SEM and not for XGBoost or SHAP.

3.2.2. Physiological Indicators

Physiological indicators like eye movements and heart rate are strongly correlated with changes in cognitive and mental workload [54]. Given the dynamic nature of driving scenarios, it is advantageous to use multiple indicators to assess drivers’ mental workload effectively. Six physiological indicators are selected: growth rate of heart rate (GRHR) and root mean square of successive differences (RMSSD) for ECG; growth rate of pupil diameter (GRPD); average fixation duration (AV FD); average saccade velocity (AV SV); and average saccade size (AV SS) for eye movement.

These metrics are proven effective for measuring the objective workload in lane-changing and scenario-switching tasks. GRHR and GRPD are calculated as the ratio of the measured increase to the baseline value, reflecting the driver’s responsiveness to external stimuli. Saccade and fixation metrics are critical for understanding visual perception and are processed using the built-in algorithms of Begaze 3.7 software. AV FD is indicative of cognitive workload, while AV SS and AV SV are frequently used in driving distraction studies to gauge the impact of secondary tasks such as lane changes or sign recognition.

3.2.3. Subjective Demand

NASA-TLX, developed by Hart and Staveland (1988) [17], is utilized to assess perceived driving load and calibrate subjective demands, including mental, temporal, and task-related aspects. This subjective measure enhances sensitivity to underload and overload situations, effectively complementing the physiological indicators collected. NASA-TLX comprises six subscales—mental demand, physical demand, temporal demand, performance, effort, and frustration. Participants rate each on a scale from 0 to 10, facilitating both independent assessments of each dimension and aggregate evaluations of combined factors. In this study, we used a version of NASA-TLX adapted for the Chinese context, which was translated and back-translated to ensure the accuracy and cultural relevance of the content, preserving the original meaning as closely as possible.

To streamline analysis within the Risk Allostasis Theory (RAT) framework, mental demand (MD) and temporal demand (TD) are treated as separate indicators, while performance, effort, and frustration are consolidated into a single metric called work/task demand (WD). This metric represents task difficulty and allows a nuanced understanding of how different scenarios and tasks influence subjective demands. This approach enhances sensitivity, effectiveness, and the depth of insights from the analysis [55].

3.2.4. Driving Performance and Conflict Indicators

To assess the stability and safety of driving performance in diverging areas, three types of indicators are selected: longitudinal, lateral, and conflict.

(1) Longitudinal Performance: This dimension evaluates the stimulus-response mechanism of car-following through metrics such as acceleration/deceleration and vehicle spacing. The selected indicators include average speed (AV Speed), standard deviation of speed (SD SP), standard deviation of acceleration (SD ACC), and distance headway (DH). These metrics are essential for assessing longitudinal control as they closely correlate with the objective risk associated with driving.

(2) Lateral Performance: For lane-changing analysis, studies typically employ measures like lane change time (LC Time) and angle (LC Angle) to assess maneuver characteristics. Additional metrics such as the standard deviation of lateral position (SDLP) and steering angle (SDSA) are also used to evaluate lane-keeping stability. These measures are employed to gauge lateral control degradation in this study.

(3) Conflict indicators: These serve as alternatives for assessing potential loss of control and crash tendencies and are extensively used in real-time risk assessments for autonomous vehicles [56,57]. To indirectly reflect the potential crash risk in diverging scenario, Modified Time to Collision (MTTC) is chosen to reflect potential rear-end risks in diverging scenarios due to its inclusion of vehicle dynamics; Deceleration Rate to Avoid a Crash (DRAC) is selected as a conventional deceleration-based index that reflects the collision kinematic condition [58,59,60].

The calculation of minimum MTTC and maximum DRAC for each trip is based on real-time data from surrounding vehicles, using the following formulas:

$$\mathrm{M}\mathrm{T}\mathrm{T}\mathrm{C}={\displaystyle \frac{-∆\mathrm{v}\pm \sqrt{∆{\mathrm{v}}^2+2∆\mathrm{a}({\mathrm{x}}_{\mathrm{l}}-{\mathrm{x}}_{\mathrm{f}}-{\mathrm{D}}_{\mathrm{l}}}}{∆\mathrm{a}}}$$

(1)

$$\mathrm{D}\mathrm{R}\mathrm{A}\mathrm{C}={\displaystyle \frac{{({\mathrm{v}}_{\mathrm{f}}-{\mathrm{v}}_{\mathrm{l}})}^2}{2({\mathrm{x}}_{\mathrm{l}}-{\mathrm{x}}_{\mathrm{f}}-{\mathrm{D}}_{\mathrm{l}})}}$$

(2)

where x_l and x_f are observed positions of the front bumpers of the leading vehicle and following vehicle; v_l and v_f are the observed speeds of the leading and following vehicles; Δv = v_f − v_l is the difference in the speeds of vehicles; Δa = a_f − a_l is the difference in their accelerations; D_l is the length of the leading vehicle.

After data preprocessing and calculation of these indicators, a basic descriptive statistical analysis of the observable variables is presented in

Table 2. Description statistics of workload and driving performance indicators.

Latent Variables	Variables	Mean	Std.	Max	Min
Objective	GRHR (%)	22.078	5.493	38.045	3.134
workload	RMSSD (ms)	22.383	3.338	31.307	14.668
	AVFD (ms)	442.090	67.990	729.609	202.229
	GRPD (%)	31.019	11.754	61.987	4.897
	AV SV (pixel/s)	1.205	0.522	2.706	0.247
	AV SS (pixel)	53.724	26.925	118.780	8.085
Subjective	MD	17.730	3.033	28	9
demand	TD	18.779	2.939	26	11
	WD	19.147	2.734	26	11
Longitudinal	AV speed (km/h)	73.970	4.194	88.940	59.127
performance	SD ACC (m/s2)	0.497	0.192	1.109	0.044
	SD speed (km/h)	7.515	2.564	14.338	1.568
	DH (m)	51.229	15.750	104.284	15.864
Lateral	LC time (s)	3.523	1.237	8.113	1.258
performance	SDSA (Degree)	14.445	4.252	28.466	3.686
	SDLP (m)	6.370	1.683	11.622	1.907
	LC angle (Degree)	16.438	5.511	28.143	5.037
Conflict	Min MTTC (s)	2.179	1.394	6.325	0.143
indicators	Max DRAC (m/s2)	1.524	1.151	5.179	0.000

. To explore the relationships between these indicators, a correlation heatmap is displayed in Figure 7.

3.3. Statistical Approach

A two-step evaluation is employed to analyze the factors contributing to driving performance, considering both observed and latent variables. Firstly, XGBoost models are developed to analyze nonlinear relationships among observational indicators. The SHAP post hoc analysis further quantifies the importance and interaction effects of these factors. The second step involves PLS-SEM to validate our theoretical framework and examine direct and mediating effects among six latent variables. Detailed descriptions of XGBoost, SHAP, and PLS-SEM methodologies are available in [26,61,62].

3.3.1. Development of XGBoost Model

XGBoost is an advanced ensemble machine learning algorithm widely used for nonlinear evaluations in transportation [61]. It provides various strengths: efficient and extendable, applicable to nonlinearity, and suitable for variable selection and post hoc analysis. XGBoost enhances model performance by aggregating multiple gradient-boosted decision trees, with each tree building upon the residuals of its predecessors. This sequential learning process can be mathematically represented as:

$${\mathrm{y}}_{\mathrm{i}}^{\left(\mathrm{t}\right)}=\sum _{\mathrm{k}=1}^{\mathrm{t}}{\mathrm{f}}_{\mathrm{k}}\left({\mathrm{x}}_{\mathrm{i}}\right)={\mathrm{y}}_{\mathrm{i}}^{(\mathrm{t}-1)}+{\mathrm{f}}_{\mathrm{t}}\left({\mathrm{x}}_{\mathrm{i}}\right){\mathrm{f}}_{\mathrm{k}}$$

(3)

where ${\mathrm{y}}_{\mathrm{i}}^{\left(\mathrm{t}\right)}$ is the final tree model evolved from the generated ${\mathrm{y}}_{\mathrm{i}}^{(\mathrm{t}-1)}$, ${\mathrm{f}}_{\mathrm{t}}\left({\mathrm{x}}_{\mathrm{i}}\right)$ is the updated tree learner, t is the number of the basic tree-based branches, and ${\mathrm{x}}_{\mathrm{i}}$ is the input set.

Former applications have proven that the performance of XGBoost models can be enhanced by 5–20% through hyperparameter tuning and cross-validation (CV) [63]. This study employs Bayesian optimization to refine the model by integrating insights from previous iterations and uses 10-fold cross-validation to ensure robustness, with the optimized models undergoing further post hoc analysis.

3.3.2. Post Hoc Analysis by SHAP

With the optimal models, SHAP is used to interpret the contribution of each factor to different dependent variables and their joint interaction effects. SHAP is a cutting-edge method that blends the principles of game theory with local machine learning predictions to decompose the output into the contributions of each input feature [22]. This approach is articulated through the equation:

$$\mathrm{F}\left(x_i\right)=\mathrm{G}\left(z_i\right)={\Phi }_0+\sum _{j=1}^m{\Phi }_{\mathit{ij}}{z}_{\mathit{ij}}$$

(4)

where $\mathrm{F}\left(x_i\right)$ is the model output; $\mathrm{G}\left(z_i\right)$ is an explanatory model for the interpretation of $\mathrm{F}\left(x_i\right)$; $z_{\mathit{ij}}\text{∈}\{0,1\}$, and when the feature j is observed, $z_{\mathit{ij}}=1$, otherwise, $z_{\mathit{ij}}=0$; ${\Phi }_0$ is the initial output without features, and ${\Phi }_i$ is the Shapley value of feature i; In the end, ${\Phi }_{\mathit{ij}}$ can be defined as the difference between the Shapley value of factor i with and without feature j, which is weighted by the sum of the marginal contribution of feature j and reflects the feature’s importance.

SHAP not only ensures consistency in interpretation across different models but also details the local impact of each feature, thereby enhancing the transparency and understandability of complex machine learning models. The ML analysis is performed based on XGBoost and shap packages under the Python 3.8 environment.

3.3.3. PLS-SEM Modeling Procedure

XGBoost-SHAP analysis offers a granular understanding of individual variable contributions to evaluation outcomes, complementing SEM, which estimates relationships among observed variables represented by latent constructs. Preliminary correlation analysis (Figure 7) indicates strong associations between several environmental and workload indicators with driving performance. To fully test the theoretical model, we employ SEM to delve into the intricate relationships among the latent variables delineated in the framework (Figure 2), including environmental stimulus, objective workload, subjective demand, longitudinal performance, horizontal performance, and conflict indicators. These variables are modeled in SEM as linear combinations of observed indicators, facilitating validation of the model constructs and assessment of the relationships’ significance and strength.

The Partial Least Square (PLS) method, a principal component-based SEM technique, was used to develop and validate the theoretical model. PLS-SEM is advantageous for our study due to its flexibility in accommodating both formative and reflective constructs and its suitability for complex theoretical frameworks. Moreover, PLS-SEM does not require data to adhere to a strict normal distribution, making it ideal for analyzing experimental results from driving scenarios [26,64].

PLS-SEM employs a two-stage approach [65]. In the first stage, the measurement model defined the relationships between observed indicators and their corresponding latent constructs. The measurement model’s reliability and validity were assessed using composite reliability (CR), average variance extracted (AVE), and discriminant validity, ensuring that each latent construct was distinct and adequately captured by the observed indicators. In the second stage, the structural model tested the hypothesized relationships among latent variables, including direct and mediated effects. Path coefficients were estimated to quantify the relationships’ strengths, and model fit was evaluated using standard indices such as standardized root mean square residual (SRMR) and normed fit index (NFI).

To further validate the model, a bootstrapping technique with 5000 resamples was used to assess the significance of path coefficients and examine mediation effects. This approach provided robust confidence intervals, allowing us to identify significant effects among the latent constructs and test the proposed hypotheses comprehensively. Additionally, R² values were computed for each endogenous construct to determine the variance explained by the independent variables.

In conclusion, the integration of XGBoost-SHAP with SEM provided a detailed examination of feature importance, enhancing the interpretability of SEM results by determining the contribution of each feature to the model outcomes. This complementary approach led to a more comprehensive understanding of how environmental and workload factors interact to influence driving behavior and conflict risk. By leveraging both methods, we were able to precisely identify key performance-related factors, unravel their interdependencies, and improve the overall interpretability of complex relationships within the model.

4. Results

4.1. Feature Importance of Individual Models

In this study, XGBoost models have been built using 10 individual indicators of driving performance and conflict risk as dependent variables. Model performance was evaluated by assessing the explanatory power of the input features, with results of cross-validation for driving indicators under different feature combinations presented in

Table 3. Model performance under different feature combinations.

Dependent Variable	External Factors		External & Workload Factors		External, Workload & Performance Factors 1
	R2	RE	R2	RE	R2	RE
LC angle	0.35	16.23	0.71	11.42	0.78	8.34
LC time	0.52	11.26	0.75	10.47	0.75	10.47
SD LP	0.28	21.26	0.65	13.11	0.68	12.78
SD SA	0.29	25.35	0.73	10.36	0.77	10.15
SD SP	0.31	23.18	0.67	11.87	0.67	11.93
SD ACC	0.18	24.46	0.55	14.12	0.61	13.21
DH	0.31	17.94	0.76	11.69	0.79	11.64
AV SP	0.23	19.34	0.36	17.24	0.35	17.24
DRAC	0.39	20.04	0.70	14.72	0.78	12.83
MTTC	0.43	20.78	0.71	13.53	0.81	12.85

¹ For the models with lateral performance as the dependent variable, the longitudinal indicators are employed for driving performance, and vice versa; All performance variables are used for DRAC and MTTC modelling.

, as indicated by the fitting criterion (R²) and relative error (RE). Findings indicate significant differences in model performance across various feature combinations. For most dependent variables, environmental information alone proved insufficient to capture the complexity of driving samples. Incorporating driving workload and performance indicators significantly enhanced the model’s evaluative capacity (R² > 0.61, RE < 13.21%), ensuring the validity of subsequent SHAP analysis. Both objective and subjective workloads performed well, likely due to strong correlations among the input variables.

SHAP analysis was further employed to quantify the specific impact of each variable on driving performance and conflict risk indicators. Figure 8 visually represents the feature importance for each dependent variable in summary plots, where (1) features are ranked by their importance (average SHAP value); (2) the color of sample points corresponds to feature values; and (3) local SHAP values indicate whether feature value changes promote or resist the dependent variable, reflecting the impact of specific factors. Due to poor fitting results (R² < 0.36), the feature importance involving AVSP is not listed.

For most dependent variables, driver workload, as reflected by RMSSD and AVSS, had the most significant impact, followed by environmental factors and driving control. The influence of environmental stimuli on MTTC and DRAC was considerably lower than other factors, indicating that driving risk is primarily determined by micro-driving states. Notably, key features affecting lateral and longitudinal behavior differed significantly; shorter connection distances increased the difficulty of lateral operations, particularly during lane-changing. In contrast, longitudinal control indicators were more sensitive to the information volume of signs, likely due to cognitive stress from information overload. Similarly, subjective workloads affecting lateral and longitudinal performance were driven by different key features, such as TD for lane-changing and MD for car-following.

The distribution of SHAP values provides an indirect illustration of the marginal effects of various explanatory variables. Among visual workload indicators, AVSS exhibited the highest feature importance, characterized by a balanced SHAP value distribution. In tunnel-interchange scenarios, frequent gaze shifts inevitably impact decision-making and increase driving risk. RMSSD is also a significant workload indicator, with its SHAP distribution showing a threshold effect impacting lateral performance.

For conflict risk, Figure 8h,i highlight the robust explanatory capability of DH and SDLP in terms of temporal proximity and kinetic characteristics. Stable lane-keeping performance effectively reduces crash risks, and a reduction in AVSS and LC angle likely elevates MTTC, thus enhancing the safety margin. Moreover, a greater headway distance and lower SD ACC provide drivers ample time to assess forward traffic conditions and execute appropriate actions, mitigating longitudinal conflict trends.

In fact, most features show limited contribution to the evaluation results, including several high-related variables. For example, the correlations between several lateral and longitudinal indicators do not enhance their mutual predictive efficacy. This could be attributed to the interaction and mediating effects of the explanatory variables, particularly due to the presence of driving workload and demands, which makes extra features insufficient in providing added information.

4.2. Main and Interaction Effects

The summary plots above quantitatively analyze the relationship between individual variables and driving performance. Figure 9 further interprets the main and interaction effects of combined variables using SHAP dependence plots. In each sub-figure, the horizontal axis represents the value of the independent variable, and the color of the points represents the value of the interaction variable. The vertical distribution of SHAP values reflects the interaction effects, providing additional information on how the two variables collectively affect the model output.

The dependence analysis focuses on three dependent variables: LC time, DH, and MTTC, representing different aspects of driving performance. Figure 9a compares the main and interaction effects of CD, TD, and AV SS on LC time, showing a threshold distribution. Results show that a lower CD (less than 300 m) leads to a significant increase in temporal demand, reducing LC time by 0.4 s to 1.2 s. While for CD above 500 m, LC duration is relatively stable and mainly depends on subjective workload. Both TD and AV SS significantly influence LC duration. Moderate levels of these variables do not notably affect driving stability, but excessive levels (TD > 20, AV SS > 75 pixels) result in a substantial decrease in lateral performance. Additionally, the interaction effects indicate that reduced spacing in tunnel-interchange sections increases drivers’ workload and potentially compromises their ability to select the correct accepted gap, thus hindering lateral control.

DH assesses drivers’ longitudinal performance concerning operation and driving safety. Unlike lateral indicators, the influence of objective workload on DH surpasses that of environmental and subjective factors. Figure 9b indicates that increasing heart rate variability causes a rapid decline in SHAP values, which stabilize between 0 and −7 as RMSSD exceeds 25, negatively impacting car-following stability. Pronounced interaction effects between RMSSD, MD, and IV are observed; lower IVs improve SHAP values under similar workloads, as illustrated in the left graph of Figure 9b. Excessive signage information can lead to sensory overload, which impairs drivers’ attention and reaction times, thereby influencing DH variability (−5 to 3 m). The threshold effect is also observed when the information amount surpasses 9 units, resulting in a sharp decrease in DH. Hence, keeping information within 9 units may enhance car-following safety.

Figure 9c illustrates the dependence between conflict risk and various factors. The threshold for CD on MTTC is higher (700 m) compared to that of LC time (500 m). When CD is below 500 m, changes in IV do not appear to significantly affect MTTC, suggesting that the external environment exerts a limited influence on conflict risk. For driving performance, the risk of conflict tends to generally decrease with increasing DH and LC time, but with variations in distribution. The MTTC growth rate decreases gradually with DH and stabilizes after 50 m, while the SHAP value for LC time has a quadratic distribution, reaching a minimum of −0.5 at 3 s. In addition, AV SS remains stable when LC time is below 3.5 s, which also supports the behavioral adaptation between driving workload and objective risk. These observations highlight the dynamic relationship between mandatory lane change characteristics and driving risk from a human factor perspective.

4.3. PLS-SEM

Based on the theoretical framework, PLS-SEM is used to investigate the factors influencing the driving performance and risk under tunnel-interchange connection scenarios. To comprehensively evaluate the performance of the PLS-SEM model,

Table 4. Overall fit indices for the SEM model.

Evaluation Metric	Definition and Criteria	Result
SRMR	Standardized Root Mean Square Residual (<0.1 indicates good model fit)	0.047
NFI	Normed Fit Index (>0.9 indicates substantial model improvement)	0.907
χ2/df	Chi-square to degrees of freedom ratio (<2 indicates good fit)	1.437
R2	Coefficient of Determination (higher values indicate stronger explanatory power)	0.647
Q2	Predictive relevance (>0.5 indicates satisfactory predictive accuracy)	0.574

presents the model’s fit and explanatory power with respect to the theoretical framework, along with an explanation of each evaluation metric and the corresponding results obtained in this study.

The results indicate that the model exhibits statistically significant advantages in terms of goodness of fit, improvement over the baseline model, model complexity, and predictive explanatory power. This evidence supports the validity of the proposed SEM model and its suitability for the theoretical framework.

4.3.1. Measurement Model

Table 5. Reliability and validity test of the measurement model.

Construct	Type	Items	Loadings	Weights	VIF	CR	CA	R2
External stimulus (ES)	Formative	TV	0.379 ***	0.334 ***	1.003	/	/	/
		CD	0.876 ***	0.852 ***	1.017	/	/	/
		IV	0.319 ***	0.300 ***	1.028	/	/	/
Objective workload (OW)	Reflective	AV SS	0.789 ***	0.377 ***	/	0.788	0.779	0.527
		GRPD	0.840 ***	0.376 ***	/
		RMSSD	0.869 ***	0.445 ***	/
Subjective workload (SW)	Reflective	MD	0.935 ***	0.380 ***	/	0.934	0.894	0.495
		TD	0.879 ***	0.317 ***	/
		WD	0.979 ***	0.374 ***	/
Lateral Performance (LAP)	Reflective	LC time−1	0.881 ***	0.456 ***	/	0.782	0.758	0.620
		SDLP	0.821 ***	0.442 ***	/
		SD angle	0.751 ***	0.312 ***	/
Longitudinal Performance (LOP)	Reflective	DH−1	0.890 ***	0.476 ***	/	0.800	0.711	0.554
		SD ACC	0.773 ***	0.318 ***	/
		SD SP	0.814 ***	0.406 ***	/
Conflict risk (CFR)	Reflective	DRAC	0.912 ***	0.585 ***	/	0.774	0.768	0.585
		MTTC−1	0.889 ***	0.525 ***	/

*** means p value of each item loading is less than 0.01; LC time, DH and MTTC are inverted to facilitate PLS modeling.

and

Table 6. Convergence validity and discriminant validity test.

Latent Variables	AVE	Discriminant Validity
		(1)	(2)	(3)	(4)	(5)	(6)
(1) Conflict indicator	0.811	0.901 *
(2) Environmental stimulus	/	0.542	/
(3) Lateral performance	0.672	0.585	0.627	0.820
(4) Longitudinal performance	0.684	0.532	0.449	0.349	0.827
(5) Mental workload	0.869	0.597	0.612	0.713	0.533	0.932
(6) Objective workload	0.694	0.662	0.653	0.705	0.566	0.698	0.833

* The bolded diagonals are the square root of AVE, and the off-diagonal elements are correlation among constructs.

present the results of the PLS component-based analysis of the measurement model. The latent variables are classified as either reflective or formative, each constructed from two or three factors refined through multiple model revisions.

Table 5 shows that the standardized factor loadings for each reflective item exceed 0.6, and the weights for each formative item are above 0.2, all with p-values below 0.01. Additionally, both the Composite Reliability (CR) and Cronbach’s Alpha (CA) for each latent variable are greater than 0.7, indicating reliable measurement and consistent item performance within each latent variable.

Table 6 showcases the Average Variance Extracted (AVE) and the relationship between the constructs. The AVE values, all exceeding the threshold of 0.50, reflect the strong connection of each construct with its indicators. The correlations among constructs demonstrate the degree of overlap, with a comparison of the square root of AVE to these correlations confirming that each construct is more strongly associated with its own measures than with others. Collectively, these results affirm the validity and reliability of the constructs and substantiate the appropriateness of the observation indicators for each latent variable.

4.3.2. Structural Model

The SEM establishes a structural model that can relate each measurement model under the hypothesis framework. Figure 10 depicts the refined structural model with path coefficients. The paths between latent variables measure the strength of the direct effects, and the indirect effects reflect the mechanism of potential mediation. Total effects can be measured by the sum of direct and indirect path coefficients.

In Figure 10, dashed lines indicate that certain relationships between latent variables are primarily influenced by mediation effects, with no significant direct effects observed. Specifically, the direct relationships between environmental factors and conflict risk, subjective workload and conflict risk, as well as environmental factors and longitudinal driving performance, were found to be non-significant. This implies that these effects are largely mediated through other variables rather than occurring directly. For instance, the influence of environmental factors on conflict risk is mediated through subjective workload and adaptive driving strategies.

The unsupported direct hypotheses indicate that the impact of environmental factors on outcomes like conflict risk and longitudinal driving performance is not straightforward and occurs through indirect pathways. Therefore, interventions that solely focus on modifying environmental conditions may not have a direct effect on reducing conflict risk. Instead, addressing mediators such as drivers’ workload (OW and SW) and adaptive behaviors (LOP and LAP) are crucial for achieving effective safety outcomes. This highlights the importance of considering mediating factors to fully understand how external conditions translate into changes in driving behavior and safety outcomes.

Table 7. Comparison of total effects among constructs.

Constructs	Effects	p Value	f2	Constructs	Effects	p Value	f2
ES-LAP	0.629	0.000	0.172	ES-CFR	0.546	0.000	0.008
OW- LAP	0.269	0.000	0.267	SW-CFR	0.145	0.033	0.010
SW- LAP	0.355	0.000	0.056	OW-CFR	0.439	0.000	0.140
ES-LOP	0.452	0.000	0.010	LAP-CFR	0.222	0.008	0.367
SW- LOP	0.195	0.021	0.131	LOP-CFR	0.236	0.000	0.174
OW- LOP	0.340	0.000	0.214

,

Table 8. Indirect effects for driving performance.

	Lateral Performance					Longitudinal Performance
	Direct Effect	Total Effect	Indirect Effect		VAF	Direct Effect	Total Effect	Indirect Effect		VAF
			SW	OW				SW	OW
TV	0.022	0.14	0.05	0.068	0.84	0.16	0.261	0.054	0.047	0.39
CD	−0.303	−0.645	−0.138	−0.204	0.53	−0.030	−0.24	−0.099	−0.111	0.88
IV	−0.151	−0.008	0.027	0.116	/	0.204	0.312	0.028	0.08	0.35
ES	0.403	0.629	0.049	0.177	0.36	0.109	0.452	0.12	0.223	0.76

and

Table 9. Indirect effects for conflict risk.

	Conflict Risk
	Direct Effect	Total Effect	Indirect Effect				VAF
			SD	OW	LAP	LOP
TV	0.074	0.224	0.005	0.053	0.005	0.037	0.67
CD	−0.069	−0.444	−0.013	−0.160	−0.069	0.012	0.84
MV	0.036	0.198	0.003	0.091	−0.034	0.047	0.82
ES	0.092	0.546	0.012	0.196	0.052	0.025	0.83
SW	0.020	0.145	/	/	0.079	0.046	0.86
OW	0.298	0.439	/	/	0.06	0.08	0.32

further show the total and indirect effects in the structural model.

Table 7 validates the association between the theoretical framework elements with a significant total effect (p-value < 0.05) for all constructs. The paths in Figure 10 show that ES factors directly influence both SW and OW, which in turn impacts LAP and LOP. However, the dashed path from ES to LOP suggests that this effect is mediated through driving workload, as it cannot be directly quantified. This underlines the intricate relationship between environmental factors and driving behaviors, emphasizing how changes in environmental stimuli dynamically interact with drivers’ cognitive and physiological states to their driving performance.

Table 8 then validates the mediating effects of SW and OW on the relationship between external stimuli and driving performance. VAF (variance accounted for) measures the strength of mediation, where VAF values of 0 to 0.25, 0.2 to 0.75, and 0.75 to 1 correspond to no, mixed, and full mediating effects. The results show a non-negligible mediating effect for all external variables, with OW dominating the mediating effects for both lateral and longitudinal operations, which illustrates its critical role in translating external information into measurable changes in driving performance, particularly in high-demand small spacing driving scenarios, where the cognitive and sensory load can be intense.

Table 9 highlights the pivotal role of driving workload and performance in shaping conflict risk, despite the potential influence of external factors. It emphasizes the significance of both direct and indirect effects in managing complex driving risks, influenced by physiological traits. Findings demonstrate that while subjective workload measures such as perceived effort and stress provide crucial descriptions about a driver’s internal state, objective measures like HRV and eye tracking furnish more direct insights into physiological impacts of driving tasks. This interaction between subjective and objective factors is essential for understanding and mitigating risks in complex driving environments.

5. Discussion

To investigate the impact of small spacing sections on driving behavior and safety, this study conducted driving simulation across 60 uniquely designed small spacing sections, collecting 19 observable variables related to driver status and performance. Using the BAT integration framework, this study addresses the complexity of driving behavior by combining multiple theoretical perspectives, offering a more holistic understanding of how drivers adapt to challenging environments. The study combined single-factor ranking (SHAP), interaction effects (dependence plots), and multi-factor analysis (SEM) to explore the key variables influencing driving behavior from multiple dimensions. These findings help to gain a better understanding of the direct or indirect relationships for driving-related factors and the behavioral mechanism under tunnel-interchange environments.

5.1. Impact of External Stimulus Factors

The spacing between the tunnel and interchange is a crucial variable affecting lateral control and conflict indicators. Our results reveal that 75% of lane change duration is distributed in 2–6 s, which is shorter than that of general sections [66]. This highlights the urgent demand for mandatory lane changes in small-spacing sections, while the shorter LC duration increases the risk of traffic conflicts. When the CD is less than 700 m, the MTTC is significantly reduced, which is higher than the threshold length of the diversion effect of 500 m [67]. This may be because drivers need 100–200 m to adapt to the ‘white hole effect’ at the tunnel exit [10], reflecting the negative impact of environmental switching on conflict risk.

Information overload on traffic signs may increase drivers’ visual cognition burden and degrade longitudinal performance. An increased density of redundant information requires drivers to slow down to ensure sufficient recognition response [14]. However, the interaction effects of IV and CD on conflict risk are not apparent when CD is smaller than 500 m. This may be due to the urgency of the LC task reducing the demand for secondary tasks such as visual recognition. To ensure sufficient reaction time, it is recommended to provide drivers with clear and concise information regarding upcoming diverging ramps and to limit the number of information units to no more than six, which is fewer than, in general, interchange sections [4].

It is important to note that the traffic flow in the simulation environment was based on field traffic surveys with a conservative service level under Level II, resulting in a weak correlation with other factors. Based on the existing literature and current results [15], it can be inferred that conflict risk may arise with increasing traffic density, which warrants further research to verify this assumption.

5.2. Behavior Mechanism in Small Spacing Sections

This study enhances the classical framework by incorporating several behavioral theories to properly reflect the conceptual subjective and objective workload and demonstrated the structural validity of the integrated latent variables through SEM. The integration of multiple theoretical models, such as BAT, RAT, and the DWT, provides a comprehensive framework for understanding driver adaptation in complex environments. This integrative approach offers new insights into the mechanisms through which environmental factors, workload, and performance interact, surpassing the explanatory capacity of individual models. The framework reveals that the “environment-performance-risk” process is typically mediated by mental demands rather than direct effects. This finding emphasizes the significant impact of environmental stimuli on workload, which ultimately determines driving performance from a human factor perspective.

High-load environments have a significant impact on drivers’ cognitive and visual abilities, with objective indicators such as AVSS and RMSSD showing greater contributions to both lateral and longitudinal control. Saccade size is commonly considered to reflect a driver’s visual strategy and attention allocation during the scanning and processing of environmental information [68]. Drivers’ average saccade amplitude exhibits an upward trend in sections with smaller connection distance, indicating an increase in their ability to perceive the surrounding environment and potential risks. However, a larger scanning range may reduce the spatial resolution of vision and lead to distraction. Our findings suggest that visual overload may occur when AVSS exceeds 75 pixels, significantly increasing the risk of lane deviation and traffic conflicts. Similarly, complex external environments increase driver anxiety and stress, making RMSSD highly sensitive to driving behavior.

Physiological indicators reflect the stress response of drivers under the influence of small spacing environments [28]. Appropriate increases in mental load are in line with the positive regulation of task demands, which is conducive to the coherent integration of peripheral traffic information, consistent with deep cognitive mechanisms such as visual search patterns and activation effects [69]. However, the construct validity test revealed relatively lower correlation between subjective and objective workload, indicating discrepancies in how drivers perceive and react to risk under multitasking conditions. This discrepancy underlines the need for a theoretical framework that can accommodate both subjective and objective measures to better capture adaptive behaviors and self-assessment biases, thereby providing a more complete understanding of the driver risk assessment process. Such biases in risk perception are recognized as a significant contributor to traffic accidents [70]. Hence, practical efforts should focus on developing driving risk adjustment strategies based on OW indicators to mitigate these biases.

Regarding driving behavior, RAT offers a framework for explaining the interaction effect between car-following spacing, LC duration, and conflict risk [50]. When drivers maintain shorter headways, they may resort to rapid lane changes to navigate through critical gaps, which can increase the aggressiveness of lateral movements. This behavior is an adaptive mechanism aimed at keeping perceived risk within a manageable range [39]. The effectiveness of this compensatory behavior is supported by the observation that there are no significant differences in MTTC intervals for medium-to-low LC durations, though the distribution is more varied. However, it is important to note that excessive aggression in driving can significantly elevate conflict risk, as indicated by the relationship between DH and MTTC shown in Figure 9c.

This observed trend of change also supports the threshold relationship between mental demands, behavior, and risk, suggesting that these factors maintain dynamic stability within a controllable range but escalate rapidly once a certain threshold is exceeded. Regarding subjective demands or risk perception, these findings align with Summala’s safety margin model [71]. The distribution of objective risk similarly concurs with experimental results from RAT studies [48,49].

Moreover, the findings support the necessity of a comprehensive BAT framework that integrates these theories to maintain a balance between safety and efficiency. By incorporating the “stimulus-cognition-response” cycle and understanding both subjective and objective aspects of workload, this study demonstrates the added value of an integrated BAT framework in addressing the unique challenges posed by complex driving environments. Consequently, the design of small-spacing sections should strive to balance safety and efficiency by integrating objective environmental factors such as connecting spacing, traffic flow, and signage with optimized management of driver temporal demands and conflict risk.

6. Limitations

This study employs driving simulation technology, which is essential for investigating the complex driving environments of tunnel and interchange connecting sections. These scenarios require high levels of experimental control to effectively isolate and analyze specific behaviors and variables. The driving simulator provides an ideal setting for such precise manipulation and repeatability of conditions, which would be challenging and potentially hazardous to replicate in real-world environments. However, while simulators ensure safety and detailed control, they inevitably lack some real-world complexities, which may lead to certain differences between driver behavior in the simulator and on actual roads, potentially affecting stress responses, decision-making, and error rates to some extent. Also, the absence of physiological or psychological stress indicators, such as electrodermal activity or electroencephalography, limits the study’s ability to capture a more detailed understanding of drivers’ stress responses in these scenarios. Although this does not invalidate our findings, it suggests that future studies could further enrich the insights into emotional and cognitive states by incorporating these measures.

Furthermore, the selection of participants may impact the applicability of our findings. Our study involved exclusively younger drivers, aged 25–33, limiting the generalizability to more experienced or older drivers. Literature suggests that experienced drivers typically exhibit quicker reaction times and more proactive driving behaviors, influencing their decision-making regarding speed control and following distances [72]. This demographic limitation raises concerns about the representativeness of other drivers.

Additionally, the within-subjects experimental design, where drivers encountered multiple scenarios in randomized orders, could also introduce biases. While randomization helps mitigate some effects, repeated exposure to the simulator may lead to increased familiarity with the experimental setup, potentially altering drivers’ behaviors over successive trials and leading to skewed results.

To enhance the generalizability of our findings, future research could incorporate field experiments to validate our results and identify additional influential factors not evident in simulator settings. Expanding the demographic diversity of participants would also broaden the applicability of the outcomes. Moreover, incorporating more sensitive physiological indicators in future studies could provide deeper insights into drivers’ emotional and stress responses, especially in high-complexity environments. To mitigate potential biases from simulator familiarity, varying the complexity of scenarios should be considered to ensure the robustness of the results.

7. Conclusions

This study employs simulation experiments to investigate driver behavior and associated risk patterns in the small-spacing section of tunnel-interchange, considering the impact of stimulus-induced driving workload. The integrated BAT framework is adopted to emphasize the direct and mediating effects of driving workload on driving performance, while SHAP and SEM analyses provided insights into the behavioral feedback mechanisms in complex environments. The following conclusions can be drawn:

(1) External factors such as smaller spacing and overloaded signage significantly increase drivers’ task difficulty, degrade driving performance, and contribute to elevated conflict risk. To ensure adequate response time, the recommended connection distance should be more than 700 m, with no more than six sign information units.

(2) High-load environments significantly impact drivers’ stress response and visual abilities, with objective indicators showing greater contributions to both lateral and longitudinal control. This highlights the dynamic relationship between driving characteristics and conflict risk from a human factor perspective.

(3) From the perspective of behavioral adaption, shorter headways in small-spacing sections necessitate quick lane changes in critical gaps, prompting drivers to adopt more aggressive lateral maneuvers to control perceived risk within an acceptable range. These can help inform the development of traffic management strategies aimed at reducing risky driving maneuvers and enhancing overall safety in small-spacing areas.

(4) Driving workload and performance are proven to dominate the fully mediating effects between external factors and conflict risk. Physiological or behavioral indicators, such as Saccade size, RMSSD, lane deviation, and headway, show high sensitivity to dynamic risk and can be used as monitoring indicators to provide effective decision support for automation applications. These indicators could be instrumental in developing advanced driver assistance systems that help monitor and mitigate high-risk driving behaviors in real-time.

It is crucial to acknowledge that small spacing sections of tunnel-interchanges impose significantly higher cognitive load and task demands on drivers compared to the general diversion area. This partially explains the higher crash rates observed in these areas and highlights the need for targeted prevention measures, such as improved signage setting, increased spacing, and tailored driver education programs. By emphasizing these safety measures, the findings of this study can contribute to reducing the likelihood of accidents in such critical areas.

Future research could benefit from extensive collision statistics and traffic conflict data, utilizing extreme value theory to explore accident mechanisms and potential risks in small-spacing sections. Additionally, by increasing the number of engineering samples, field driving tests could be conducted to assess the impact of real-world factors, such as dynamic traffic characteristics and weather conditions.