The Mediating Role of Inattentional Blindness Between Risk Propensity and Risk Perception: An Eye-Tracking Study in Confined Spaces

Abstract

Confined space operations are characterized by environmental complexity and latent hazards, where failures in human risk perception represent a primary precursor to industrial accidents, posing a significant challenge to sustainable occupational health and safety (OHS) management. This study investigates the mechanism by which individual traits (risk propensity) influence risk perception performance through cognitive processes (inattentional blindness). Utilizing psychological scales and eye-tracking technology, we quantitatively analyzed the visual search behaviors of participants with varying risk propensities across typical confined space hazard scenarios. The results indicate that individuals with high risk propensity tend to adopt a “random-exploratory superficial scanning strategy,” characterized by significantly delayed Time to First Fixation (TFF) and lower Fixation Count (FC) within critical hazard areas compared to the low-risk propensity group. Statistical analysis reveals that inattentional blindness exerts a full mediating effect between risk propensity and risk perception performance, accounting for 72.56% of perception failures. This research confirms that an imbalance in attentional resource allocation leads to higher cognitive omission of salient hazards among high-risk propensity individuals. These findings provide a theoretical foundation for cognitive reliability assessment and the design of sustainable safety training programs in high-risk industries, ultimately contributing to the social sustainability and well-being of industrial workforces.

1. Introduction

Confined space operations consistently rank among the most fatal operational types in industrial production due to enclosed environments, poor ventilation, and latent hazards (e.g., toxic gases, flammable and explosive substances). Among these, underground mines represent the most typical and complex confined space environments. Identifying and preventing hazards such as gas explosions and toxic gas accumulations in mines is crucial for sustainable occupational health and safety (OHS). Traditionally, workplace risk assessment relies on established standards and systematic methodologies, such as the hazard identification and risk matrix approaches out-lined in ISO 45001 or specific national occupational safety guidelines. These conventional methods primarily evaluate the probability and severity of potential incidents related to physical and chemical hazards, forming the baseline for safety protocols. However, they often fall short in accounting for the dynamic cognitive failures of individual operators under stress.

Recent literature emphasizes the critical need to investigate the interplay between individual safety attitudes and systemic resilience across diverse high-risk operational contexts. For instance, recent research evaluating safety climate in aviation highlights that positive organizational perceptions and individual safety commitments do not automatically translate into adaptive resilience during unexpected events unless sup-ported by targeted training mechanisms [1]. Building on this broader occupational safe-ty motivation, our study focuses on confined space operations to understand how cognitive antecedents—specifically, how individual risk propensity influences hazard perception through the bottleneck of inattentional blindness. Understanding these mechanisms is paramount for developing resilient, cognitive-based safety interventions and aligning industrial practices with sustainable worker well-being goals.

Despite substantial industrial investments in equipment monitoring and safety protocols, the incidence of confined space accidents remains high. Despite substantial industrial investments in equipment monitoring and safety protocols, the incidence of confined space accidents remains high. Accident investigations indicate that human error is a core precursor to such tragedies, with recent empirical studies highlighting that hazard perception failures directly trigger unsafe behaviors [2]. Operators frequently face extremely high cognitive loads, particularly in complex operational environments, which significantly exacerbates cognitive distraction and diminishes safety performance [3]. When primary cognitive resources are allocated to core tasks (e.g., pipeline welding, equipment maintenance), operators are highly prone to a “looking but not seeing” phenomenon regarding peripheral safety warnings and potential hazards. This deep-seated cognitive omission directly causes failures in risk perception. Similar to the medical field, where poor interface legibility can lead to fatalities during ventilator operations [4], cognitive omissions in confined space operations are often driven by inefficient information processing under high-pressure environments.

In cognitive psychology, this phenomenon is termed inattentional blindness (IB). It refers to the inability of an observer to perceive an otherwise salient, unexpected stimulus in the visual field when attention is intensely focused on a specific task [5]. Fockert et al. noted that IB profoundly reflects the inherent limitations of human attentional resources [6]. In the practical context of confined space operations, perceiving potential hazards is often unconsciously downgraded to a secondary task by operators. When the perceptual load of the primary task is excessively high, limitations in working memory capacity lead to failures in selective attention [7]. Consequently, IB frequently occurs in high-risk industrial sites, acting as a critical cognitive bottleneck that disrupts the pathway from hazard stimulus input to safe decision output, a phenomenon increasingly recognized in recent human-centered interface and assembly task designs [8].

It is crucial to theoretically distinguish between classic inattentional blindness and the phenomena observed in this study. According to the foundational framework by Mack and Rock [9], classic IB occurs when an observer fails to notice an unexpected stimulus because their attention is fully engaged in a different primary task. In the current experimental paradigm, participants were explicitly instructed to identify potential risks. Consequently, missing a hazard constitutes a visual search failure or “task-induced hazard omission” (frequently termed the “looked-but-failed-to-see” effect), rather than classic IB in its strictest sense. However, both phenomena share the same fundamental cognitive mechanisms: the depletion of attentional resources and working memory bottlenecks under high load. To align with the broader occupational safety literature, this study conceptualizes these task-induced omissions under the umbrella framework of “inattentional blindness,” while explicitly acknowledging this operational distinction.

Not all operators experience inattentional blindness under equivalent cognitive loads. Existing research demonstrates significant trait differences among individuals in their capacity to capture unexpected stimuli [10,11]. In safety management and behavioral sciences, risk propensity serves as a core variable explaining these individual differences. Risk propensity reflects not only an individual’s subjective attitude towards uncertain outcomes [12] but also represents a stable personality and cognitive trait [13]. Individuals with high risk propensity exhibit a higher willingness to take risks during decision-making. This trait influences final behavioral choices and reshapes visual search strategies during the early stages of information processing. Contextual interaction theory posits that risk propensity modulates the allocation of attentional resources through a dynamic system of “contextual features–cognitive appraisal–behavioral output” [14]. Therefore, risk propensity is highly likely a profound antecedent determining whether an individual triggers inattentional blindness in complex environments. Although risk propensity and inattentional blindness have been extensively studied in their respective fields, the current literature presents two notable limitations. First, traditional risk propensity research relies heavily on stated preference methods (e.g., economic scales) or self-report questionnaires, which struggle to objectively and precisely capture the instantaneous attentional allocation mechanisms of operators in realistic high-risk situations. Second, few studies have integrated individual traits (risk propensity), cognitive errors (inattentional blindness), and outcomes (risk perception ability) into a unified theoretical framework. Specifically, the multiple interactions and mediating effects among these variables remain unclear within the highly specific and hazardous industrial context of confined spaces.

To address this research gap, the present study focuses on confined space operations and constructs a mediation model of “risk propensity–inattentional blindness–risk perception”. Specifically, this study makes three primary contributions: (1) It innovatively integrates risk propensity, inattentional blindness, and risk perception into a unified cognitive framework within the highly specific and hazardous industrial context. (2) It transitions from traditional subjective self-reports to objective physiological measurements by utilizing eye-tracking technology to precisely quantify operators’ visual search strategies. (3) It statistically verifies the critical mediating role of inattentional blindness, providing empirical evidence that an imbalance in cognitive resource allocation directly leads to safety perception failures.

Based on these objectives, we propose the following hypotheses:

H1. 

Individual risk propensity significantly influences visual attention allocation and search strategies (reflected in eye-tracking metrics) when identifying hazards in confined spaces.

H2. 

Inattentional blindness (IB) serves as a critical mediator in the relationship between risk propensity and risk perception performance.

Rather than presenting a direct operational safety guideline for experienced field workers, this research is strictly positioned as an exploration of fundamental cognitive mechanisms, reflecting how attentional allocation is modulated by risk traits under simulated risk conditions.

2. Materials and Methods

2.1. Participants and Risk Propensity Assessment

The worker population in confined space operations is highly diverse, exhibiting significant variations in professional characteristics, work environments, safety training, demographics (e.g., age structure, years of experience), and safety cognitive levels (e.g., risk perception acuity, emergency decision-making efficiency). Furthermore, due to the severe physical constraints of confined spaces, such as narrow entrances and strict explosion-proof requirements, deploying eye-tracking equipment in real-world operational scenarios was practically unfeasible. Therefore, to control for confounding occupational variables and ensure the stability of participant characteristics [15], 47 undergraduate and graduate students majoring in Safety Engineering and Management Science and Engineering were recruited for this study [16]. All participants volunteered, had normal or corrected-to-normal vision, and reported no color blindness or weakness. None had prior experience with similar experiments, and all provided written informed consent approved by the university’s Institutional Review Board.

Participants’ risk propensity was assessed using the General Risk Propensity (GRP) scale [12]. This 8-item scale measures an individual’s general tendency to take or avoid risks across various situations (e.g., “I like to take risks, even if I might fail”; refer to

Table A1. General Risk Propensity (GRP) scale.

Risk Tendency Measurement Items	Strongly Disagree	Disagree	Neutral	Agree	Strongly Agree
1. I like to take chances, although I may fail.	☐	☐	☐	☐	☐
2. Although a new thing has a high promise of reward, I do not want to be the first one who tries it. I would rather wait until it has been tested and proven before I try it.	☐	☐	☐	☐	☐
3. When I have to make a decision for which the consequence is not clear, I like to go with the safer option although it may yield limited rewards.	☐	☐	☐	☐	☐
4. I like to try new things, knowing well that some of them will disappoint me.	☐	☐	☐	☐	☐
5. To earn greater rewards, I am willing to take higher risks.	☐	☐	☐	☐	☐
6. I prefer to use tested and proven methods, although a new method might turn out to be better in the future.	☐	☐	☐	☐	☐
7. I will only execute a plan when I am certain it will work.	☐	☐	☐	☐	☐
8. I seek new experiences even if their outcomes may be risky.	☐	☐	☐	☐	☐

). Responses were recorded on a 5-point Likert scale, with four items reverse-coded (Items 2, 3, 6, 7).

Following data quality screening, 6 invalid samples (due to eye-tracking calibration failures or data anomalies during the experiment) were excluded, yielding a final valid sample of 41 participants. Reliability and validity analyses of the risk propensity questionnaire demonstrated robust psychometric properties, with both Cronbach’s α and the Kaiser-Meyer-Olkin (KMO) measure exceeding 0.7, p = 0.013 < 0.05. Using a median-split threshold score of 3, the final sample was dichotomized into two categories: individuals scoring below 3 were classified as the low-risk propensity group (n = 20), while those scoring 3 or above were assigned to the high-risk propensity group (n = 21).

Although treating risk propensity as a continuous variable preserves the full spectrum of data variance, a median-split approach was specifically employed in this study to establish two distinct categorical cohorts (high-risk vs. low-risk). This dichotomization is essential for eye-tracking research, as it allows for the generation of comparative visual heatmaps and scan paths, and facilitates the execution of independent samples t-tests to identify significant differences in specific oculomotor metrics (e.g., TFF, FFD) between different risk traits.

2.2. Experimental Equipment

Participants’ eye movement data were acquired using the Tobii Pro Glasses 2, a wearable eye tracker. The device operates at a sampling frequency of 100 Hz and records real-time ocular movements utilizing corneal reflection, binocular tracking, and dark pupil methodologies. The eye-tracking apparatus was interfaced with two computer monitors: one functioned as the experimenter’s console to dynamically monitor the ongoing procedure, while the other, a 27-inch display (illustrated in Figure 1), was dedicated to presenting the experimental stimuli. Both monitors featured a standard resolution of 1920 × 1080 pixels with a refresh rate of 60 Hz.

The experimental tasks were programmed and administered using Experiment Builder software v2.5.1. This involved configuring the initial and final instruction prompts, standardizing the size and spatial positioning of the visual stimuli, defining the presentation sequence, and setting the inter-stimulus intervals.

Raw eye-tracking data were preprocessed and analyzed utilizing Ergo LAB 3.0 software. During the experiment, core oculomotor behaviors—including fixation distribution, blink frequency and saccadic trajectories—were synchronously recorded at a sampling frequency of 50 Hz.

2.3. Experimental Materials

The experimental stimuli comprised images of confined space operations obtained through field surveys and virtual simulations. An initial pool of 169 photographs capturing the original states of typical confined space accidents was established. In accordance with the Safety Regulations for Hypoxic Hazardous Operations [17] and the Safety Operation Specifications for Confined Space Operations [18], the hazards were classified into five risk categories: poisoning, asphyxiation, gas explosion, falling from height, and struck by objects (defined in

Table 1. Definition of risk types for limited space operations.

Risk Category	Definition
Poisoning	Poisoning risks caused by the accumulation of toxic and harmful gases.
Asphyxiation	Hypoxia-induced asphyxiation due to poor ventilation and carbon dioxide accumulation in underground chambers and biogas digesters.
Gas explosion	Explosion and combustion risks triggered by the accumulation of flammable and explosive gases.
Falling from height	Falls occurring during climbing operations or near unprotected openings.
Struck by objects	Atypical but highly frequent physical hazards (e.g., falling debris or materials).

). From this initial pool of 169 photos, 24 high-quality images covering these five risk categories were selected as the final stimuli after expert review and pilot testing. The selection was conducted through a rigorous expert review process. A panel comprising three safety engineering experts evaluated the images based on visual clarity, the representativeness of the specific hazards, and ecological validity to ensure they accurately reflected real-world confined space environments. Furthermore, to guarantee the objectivity and reliability of the Area of Interest (AOI) annotations, the boundaries were independently delineated by two researchers. Any discrepancies were resolved through discussion and cross-validation with a third senior researcher, ensuring a consistent and precise quantification of visual attention across all stimuli.

Subsequently, Areas of Interest (AOIs) were manually delineated for each selected image. An AOI is a specific object or spatial boundary defined by the researcher within an image based on the experimental objectives; all eye-tracking metrics are extracted and quantified based on these regions. By comparing participants’ visual behavioral patterns within the AOIs, their perception characteristics regarding potential hazard sources in confined spaces can be systematically evaluated. Aligning with the aforementioned risk categories, a total of 26 fine-grained AOIs were annotated. The specific AOI numbering and detailed risk descriptions are presented in

Table 2. AOI Numbering and Risk Information Description.

AOI No.	Specific Risk Description	AOI No.	Specific Risk Description
1	Missing guardrails (fall risk)	14	Improper safety rope setup
2	Missing blind flange for energy isolation	15	Excessive combustible gas (toluene) concentration
3	Open manhole cover (fall risk)	16	Smoking in an environment with combustible gases
4	Metal debris on the ground	17	Missing respirator
5	Stones and debris at the wellhead	18	Excessive carbon dioxide in the sewer (asphyxiation risk)
6	Uncovered manhole opening	19	Overvoltage of lighting fixtures during pipeline operation
7	Entering well without safety rope and respirator	20	Hot work inside confined equipment without mechanical ventilation
8	Illegal welding operation in basement	21	Excessive hydrogen cyanide concentration in pipeline
9	Illegal hot work in basement without clearing surrounding combustibles	22	Carbon monoxide leak due to broken pipeline in maintenance room
10	Illegal cutting operation in basement	23	Welding in an environment with explosive gases
11	Desilting sewage well without respirator	24	Low oxygen concentration in sewage pool
12	Excessive hydrogen sulfide concentration in sewage well	25	Explosive gas (hydrogen) inside pressure vessel
13	Missing respirator	26	Waterproof coating volatilizing combustible gas during basement coating operation

.

It is important to acknowledge that in real-world settings, the likelihood of a hazard resulting in an accident is dynamically compounded by the interaction between human factors, such as an employee’s age, and environmental stressors like airborne dust, inadequate lighting, and intense vibrations. For instance, older workers might possess richer experience but may exhibit delayed reaction times or decreased sensory acuity when exposed to poor lighting or high-vibration microenvironments. To rigorously control these confounding physical and demographic variables and isolate the purely cognitive impact of risk propensity, the visual stimuli selected for this study were standardized to ensure consistent optimal lighting and clear visibility, and the participant cohort was restricted to a specific age range.

2.4. Experimental Tasks and Measures

2.4.1. Phase I: Risk Perception Eye-Tracking Experiment

Prior to the formal trials, participants were thoroughly briefed on the experimental objectives, structural procedures, and potential risks in accordance with ethical guidelines. They were explicitly assured of their data confidentiality and their right to withdraw from the study at any time without penalty. Following the informed consent process, participants completed the risk propensity scale. Key terminologies and concepts within the scale were explicitly defined to facilitate accurate comprehension and ensure the validity of the responses. Upon collection of the questionnaires, participants received standardized instructions regarding the eye-tracking protocols, operational requirements, and precautions, followed by a standard calibration procedure for the eye-tracking apparatus to guarantee measurement accuracy.

A standard single-point calibration procedure was executed using the Tobii controller. Among the initial recruited cohort, 6 participants failed the calibration or required excessive recalibration attempts due to severe myopia, astigmatism, or excessive ocular reflections. Consequently, these 6 individuals were excluded to ensure high gaze-tracking ratios, yielding the final valid sample of 41 participants.

Upon initiation, an instruction screen was presented, directing participants to identify potential risks and latent hazards within the subsequent images. Following a mouse click by the participant, the visual stimuli were displayed. Each image was presented for a fixed duration of 20 s, and backward navigation was strictly prohibited. To minimize visual carryover effects, an inter-stimulus interval consisting of a 3-s blank screen was implemented, followed by a 1-s prompt screen preceding the next trial. This procedural sequence was repeated iteratively until all stimuli were exhausted. The formal experimental procedure is illustrated in Figure 2.

During this phase, to objectively evaluate participants’ unconscious perception of latent hazards, four core eye-tracking metrics [4,19] based on the predefined Areas of Interest (AOIs) [20,21] were extracted utilizing ErgoLAB 3.0 software:

• Time to First Fixation (TFF): The time elapsed (in seconds) before the gaze initially enters a specific AOI. This metric reflects the efficiency of visual attention capture by the hazard stimulus and the individual’s level of vigilance;
• First Fixation Duration (FFD): The duration (in seconds) of the initial fixation upon entering an AOI. It measures the difficulty of initial information extraction and the depth of shallow cognitive processing;
• Total Fixation Duration (TFD): The cumulative time (in seconds) of all fixations falling within a specific AOI throughout the entire exploration period. This represents the overall cognitive resources allocated to that specific hazard source;
• Fixation Count (FC): The total number of times the gaze enters a specific AOI. A higher count indicates a greater degree of repeated visual confirmation and high-frequency information resampling within that area.

2.4.2. Phase II: Impact of Inattentional Blindness on Risk Perception

In this phase, risk propensity served as the independent variable, measured via the initial questionnaire, while inattentional blindness and risk perception ability were designated as dependent variables. To further investigate the interrelationships among these variables, a supplementary Situation Awareness (SA) test was conducted following the primary eye-tracking task. From the initial cohort of 41 valid participants, a subset of 30 individuals—specifically those exhibiting abbreviated fixation durations or an absence of fixations within the predefined AOIs during Phase I—were selected for this follow-up assessment.

Specifically, these 30 individuals were selected because they exhibited abbreviated Total Fixation Durations (TFD < 200 ms, generally considered the threshold for conscious cognitive processing) or a complete absence of fixations on at least one of the predefined AOIs during Phase I. This threshold criterion was deliberately established to precisely target instances of potential cognitive omission without introducing selection bias, ensuring that the SA test specifically evaluated whether objective visual deficits directly translated to subjective perception failures.

During the SA test, each selected participant was required to re-examine the experimental images, which were now explicitly annotated with the 26 predefined hazard points. Participants were asked to retrospectively confirm whether they had noticed these specific areas during the initial free-viewing task and whether they had accurately interpreted them as safety risks. An exemplary SA test scenario is illustrated in Figure 3, depicting two specific hazards: a missing respirator and an improperly rigged safety rope. Participants provided binary responses, utilizing an “X” to denote a failure to recognize the risk and an “O” to signify that the hazard was both visually captured and accurately perceived.

To verify whether participants recognized the predefined AOIs, a relatively low or zero Fixation Count (FC) on specific images was operationalized as a failure to notice the hazard during the Situation Awareness (SA) assessment. Consequently, an inability to accurately detect a hazard inherently results in a Failure of Safety Risk Perception (FSRP). Therefore, risk perception performance was solely evaluated based on hazards that had been successfully visually captured. Figure 4 delineates the operational measurement of Inattentional Blindness (IB). Specifically, FSRP manifests in two distinct scenarios: (1) The participant fails to identify the hazard during the primary visual search task. In the current study, this phenomenon is defined as IB (i.e., failure of hazard identification) and is denoted by the parameter a. (2) The participant visually captures the hazard during the primary task but subsequently fails to cognitively appraise it as a legitimate safety risk, which is denoted by the parameter b.

The performance score for each participant’s Risk Perception Ability (RPA) was calculated using the following equation:

$$RPA={\displaystyle \frac{c}{a+b+c}}$$

(1)

where a + b + c represents the total number of hazards present in the scenario, and c denotes the number of accurately perceived hazards.

Specifically, parameter c indicates that the participant exhibited normative fixation durations and fixation counts within the relevant AOIs during the initial eye-tracking task, and subsequently accurately recognized the hazard information during the SA test. Consequently, a higher RPA score reflects a superior risk perception ability. Conversely, a higher value for parameter a (failure in hazard identification) signifies a stronger propensity for the participant to exhibit inattentional blindness.

3. Results

3.1. Eye-Tracking Metrics Across Risk Propensities

3.1.1. Temporal and Frequency Metrics

The eye-tracking metrics associated with the Areas of Interest (AOIs) quantify the extent of participants’ attentional allocation toward hazard stimuli. These include five core parameters: Time to First Fixation (TFF), First Fixation Duration (FFD), Total Fixation Duration (TFD), Fixation Count (FC), and the first fixation sequence. A comparative analysis between the high-risk propensity cohort (Group A) and the low-risk propensity cohort (Group B) reveals distinct behavioral disparities in risk perception performance within the operational scenarios. The detailed statistical results are presented in

Table 3. Statistics of the differences in the fixation-related indicators of different risk stimuli.

AOI No.	TFF(s)		Sequence No.		FFD (s)		TFD (s)		FC
	A	B	A	B	A	B	A	B	A	B
1	6.07	4.64	12	12	0.26	0.29	1.35	1.30	4.69	6.74
2	7.71	3.4	13	9	0.30	0.22	1.05	0.90	4.12	3.40
3	2.88	3.01	6	7	0.28	0.2	2.16	1.40	7.32	9.85
4	2.44	3.95	5	10	0.29	0.19	1.21	1.15	4.21	5.61
5	1.12	0.75	2	3	0.55	0.4	2.67	2.86	8.12	10.27
6	2.26	1.18	5	4	0.46	0.28	2.88	2.87	9.59	10.48
7	3.56	1.85	8	5	0.48	0.26	1.60	1.53	6.07	7.66
8	4.63	6.54	12	16	0.3	0.21	0.30	1.06	1.75	3.04
9	2.47	1.59	6	5	0.22	0.22	1.24	1.60	4.75	6.54
10	4.81	7.17	11	17	0.38	0.28	0.40	0.61	1.06	2.23
11	2.22	2.45	5	6	0.81	0.39	2.72	2.82	5.49	7.17
12	4.33	6.61	9	15	0.20	0.22	0.54	0.80	2.68	3.44
13	2.46	2.57	5	6	0.28	0.18	0.85	1.5	8.19	6.26
14	5.60	3.44	10	8	0.19	0.49	1.47	1.82	3.13	4.47
15	4.09	4.01	9	9	0.65	0.51	1.18	0.86	2.93	2.25
16	6.73	5.23	12	11	0.45	0.41	0.76	0.56	2.30	1.53
17	7.99	3.11	12	7	0.33	0.21	0.69	0.75	2.75	2.82
18	3.61	3.17	7	7	0.58	0.98	1.86	1.99	2.83	3.32
19	1.98	0.75	3	3	0.20	0.15	2.75	4.71	12.31	18.19
20	4.20	3.71	6	10	0.22	0.24	1.28	1.06	4.23	3.45
21	4.37	2.34	10	6	0.27	0.20	1.13	1.79	3.39	6.79
22	2.30	5.60	5	10	0.48	0.30	1.00	0.89	3.26	3.33
23	8.96	6.67	16	13	0.45	0.28	0.40	0.65	0.77	0.76
24	4.64	3.08	7	5	0.33	0.26	0.59	0.70	2.33	3.80
25	2.59	2.82	5	6	0.30	0.23	1.11	1.48	4.91	5.57
26	4.26	2.95	8	7	0.37	0.28	0.90	1.49	4.40	3.23

.

As illustrated in Figure 5, regarding the Time to First Fixation (TFF), the two cohorts exhibited distinct disparities in visual capture efficiency across the predefined Areas of Interest (AOIs). Specifically, for hazards such as falling from height (H1), missing blind flanges for energy isolation (H2), missing respirators (H17), and gas explosions (H23), the high-risk propensity group demonstrated significantly longer TFFs compared to the low-risk propensity group. Conversely, when encountering illegal hot work operations (H8, H10), the low-risk propensity group exhibited significantly longer TFFs. Furthermore, the marginal mean differences in TFF for H3 and H15 indicate no significant variance in the initial visual attention duration between the two groups.

The sequence number of the first fixation on an AOI represents the total number of fixations occurring before the participant’s gaze initially enters the predefined area. Analyzing this sequence can elucidate the participants’ cognitive abilities and their early sensitivity to latent hazards. As illustrated in Figure 6, regarding the hazard of “metal debris on the ground,” the high-risk propensity cohort demonstrated superior risk perception capabilities. Conversely, for the “missing respirator” hazard, the low-risk propensity cohort exhibited a higher level of cognitive processing.

Furthermore, the First Fixation Duration (FFD) reflects the duration of the initial gaze remaining within the AOI. As a critical metric for early attentional allocation, a combination of a short Time to First Fixation (TFF) and a prolonged FFD signifies that the participant has allocated a deeper level of visual processing and stronger attentional resources to that specific area. Data indicates that when encountering AOI 10, the high-risk propensity group exhibited a longer fixation duration; however, for AOIs such as 26, 7, and 14, their duration decreased to a lower level. In contrast, the low-risk propensity group reached a peak FFD of 0.98 s when facing AOI 18 (the maximum value across all data points), while their fixation durations for AOIs 3, 4, 17, 19, and 20 were relatively short. Overall, the FFDs for both cohorts were under 1 s, with the low-risk propensity group highly concentrated within the extremely low range of 0.15 to 0.3 s.

3.1.2. Visual Attention Distribution

Based on the oculomotor data acquired via eye-tracking technology, this study employed visual analysis methods to conduct an in-depth investigation of participants’ visual search behaviors. Within the paradigm of eye-tracking data visualization, heatmaps and scanpaths serve as two prototypical representational modalities, providing analytical insights from the dual dimensions of collective attentional distribution and individual cognitive chronometry, respectively. By integrating quantitative statistical analyses with visualization techniques, this study constructed a multidimensional framework for eye-tracking data analysis. Through the comparison of spatiotemporal pattern disparities in gaze trajectories, we elucidated the characteristics of visual search strategies among cohorts with varying risk perceptions.

Addressing the heterogeneous risk propensity cohorts, typical samples from the low-risk propensity participants and the high-risk propensity participants (Figure 7, Figure 8 and Figure 9) were selected for gaze trajectory visualization analysis. The detailed analytical results are presented in

Table 4. Visual analysis of eye movement trajectory map of different risk tendency groups.

Stimulus No.	Low-Risk Propensity Cohort/Individual	High-Risk Propensity Cohort/Individual
Stimulus 1	Figure 7a	Figure 7b
Stimulus 3	Figure 8a	Figure 8b
Stimulus 9	Figure 9a	Figure 9b

.

By comparing the gaze trajectories for Stimulus 1 (oil storage tank) and Stimulus 3 (open manhole cover), it was revealed that the two cohorts exhibited starkly contrasting visual search patterns. The low-risk propensity group demonstrated a salient “goal-oriented” strategy: their fixations were highly concentrated on explicit hazard zones (e.g., the bottom of the storage tank and the periphery of the wellhead).

In contrast, the high-risk propensity group displayed a typical “exploratory” strategy: their fixation distribution was highly dispersed, broadly covering the upper sections and background areas of the scenes, characterized by complex trajectories and frequent saccadic jumps.

Stimulus 9 contains two specific hazards: a missing respirator and an improperly rigged safety rope. A comparison of the gaze trajectories between the two types of individuals reveals that low-risk propensity individuals generated more regressions (revisits) during the search process, yielding a higher number of fixations within the target AOIs. Conversely, while the search trajectories of high-risk propensity individuals appeared more continuous, they critically lacked sufficient sustained attention toward the identified hazards.

Taking the aforementioned stimuli as representative examples, the Fixation Counts (FC) within the predefined AOIs were compared between the two cohorts to analyze the influence of varying risk propensity traits. The detailed analytical results are illustrated in Figure 10.

Comparative analysis reveals that during the identification of specific hazard factors (H1, H3, and H14) within the images, the high-risk propensity cohort lacked sustained visual attention on the corresponding Areas of Interest (AOIs). Their frequency of visual fixations within these hazard zones was relatively low. However, within the H13 hazard zone, the high-risk propensity group exhibited a higher fixation count. This anomalous performance may be largely driven by individual subjective preferences, resulting in a risk perception process that lacks typical identifiable patterns. Specifically, H13 (missing respirator) is located directly on the worker’s facial area. The high fixation count from the high-risk propensity group may be attributed to a strong “bottom-up” visual attraction to human faces, causing them to repeatedly glance at the area. However, their corresponding brief fixation durations suggest that this repetitive glancing was merely superficial scanning driven by stimulus salience, rather than a deep, “top-down” cognitive appraisal of the safety risk.

The heatmap is a commonly utilized two-dimensional data visualization method in eye-tracking experimental research. To conduct a visual analysis of the fixation heatmaps, complete datasets from typical samples representing different risk propensity traits (Figure 11, Figure 12 and Figure 13) were employed. By superimposing the fixation data layer by layer, the resulting heatmap comparisons are presented in

Table 5. Eye movement experiment fixation hot spot map.

Stimulus No.	Low-Risk Propensity Cohort/Individual	High-Risk Propensity Cohort/Individual
Stimulus 7	Figure 11a	Figure 11b
Stimulus 18	Figure 12a	Figure 12b
Stimulus 24	Figure 13a	Figure 13b

.

Specifically for the hazards present in Stimulus 7 (illegal welding in a basement, failure to clear combustibles such as synergists around the hot work area, and illegal cutting operations), Stimulus 18 (illegal welding within confined equipment and missing blind flanges), and Stimulus 24 (painting operations in a basement with the risk of flammable gas volatilization from waterproof coatings), the visualization results (heatmaps and scanpaths) consistently demonstrated two distinct strategies. The low-risk propensity group exhibited a selective and intensive processing pattern, characterized by visual attention highly concentrated on task-relevant hazard areas with longer individual fixation durations. Conversely, the high-risk propensity group displayed an extensive but superficial scanning strategy, where gaze points were widely dispersed across the entire scene but lacked sufficient dwell time on critical safety information.

Taking the aforementioned stimuli as representative examples, the mean Total Fixation Durations (TFD) within the predefined AOIs were compared between the cohorts with different risk propensity traits. The analytical results are illustrated in Figure 14.

Regarding fixation duration, the low-risk propensity cohort exhibited longer Total Fixation Durations (TFD) for four AOIs: H8, H9, H10, and H26. Overall, the TFD of the high-risk propensity cohort on hazard zones was lower than that of the low-risk propensity cohort; however, for hazard zones H2 and H20, their TFD was higher than that of the low-risk cohort.

The inherent characteristics and potential consequences of different hazard types vary, as do their visual representations within the images. Because each image typically contains multiple distinct hazard zones, participants may develop differentiated visual search strategies when perceiving the overall hazards embedded in the scene. Given the aforementioned disparities between the high- and low-risk propensity cohorts across four eye-tracking metrics—Time to First Fixation (TFF), First Fixation Duration (FFD), Total Fixation Duration (TFD), and Fixation Count (FC)—it is necessary to discuss the impact of individual risk propensity on eye-tracking metrics categorized by hazard type.

To further explore the interaction between risk traits and task contexts, this study categorized the 26 Areas of Interest (AOIs) into five classifications based on their hazard attributes: poisoning, asphyxiation, flammable/explosive gas, falling from height, and striking by objects (detailed in

Table 6. Classification of AOIs by Hazard Type.

Hazard Type	AOI No.
Poison	H7, H11, H12, H13, H15, H17, H21
Asphyxiation	H18, H20, H22, H24
Gas explosion	H2, H8, H9, H10, H16, H19, H23, H25, H26
Falling from height	H1, H3, H6, H14
Struck by objects	H4, H5

). This classification aims to eliminate confounding effects introduced by heterogeneous hazard types and facilitate a comparative analysis. All eye-tracking metrics were averaged according to their respective hazard categories and subjected to both parametric and non-parametric tests. Based on these analyses, we determined whether the differences in risk perception capabilities between high-risk and low-risk propensity individuals were statistically significant.

The normality of the eye-tracking metrics was assessed using the Shapiro-Wilk test via SPSS 26.0 software. For all tests, the null hypothesis posited that the data followed a normal distribution. The results are presented in

Table 7. Normality test table of eye movement characteristic indexes of different risk types.

Eye-Tracking Metrics	Hazard Type	Sample Size (N)	Median	Mean	Shapiro–Wilk (W)	p
TFF	Poisoning	41	3.644	3.863	0.955	0.236
	Asphyxiation	41	3.455	3.637	0.955	0.227
	Gas explosion	41	4.001	4.609	0.76	0.000
	Falling from height	41	3.115	4.078	0.85	0.001
	Struck by objects	41	2.64	3.024	0.859	0.001
FFD	Poisoning	41	0.281	0.338	0.81	0.000
	Asphyxiation	41	0.367	0.415	0.896	0.007
	Gas explosion	41	0.259	0.27	0.97	0.531
	Falling from height	41	0.218	0.251	0.825	0.000
	Struck by objects	41	0.19	0.335	0.793	0.000
TFD	Poisoning	41	1.634	1.565	0.967	0.453
	Asphyxiation	41	1.207	1.336	0.9	0.009
	Gas explosion	41	1.451	1.457	0.973	0.627
	Falling from height	41	1.504	1.938	0.898	0.007
	Struck by objects	41	1.945	1.985	0.942	0.104
FC	Poisoning	41	4.714	4.828	0.981	0.843
	Asphyxiation	41	3.5	3.367	0.951	0.175
	Gas explosion	41	5.083	5.062	0.968	0.475
	Falling from height	41	6.25	6.333	0.97	0.55
	Struck by objects	41	7	6.717	0.983	0.908

.

As indicated by the normality test results in Table 7, at a significance level of 0.05, the following eye-tracking metric datasets followed a normal distribution: Time to First Fixation (TFF) for poisoning and asphyxiation; First Fixation Duration (FFD) for flammable/explosive gas; Total Fixation Duration (TFD) for poisoning, flammable/explosive gas, and struck by objects; and Fixation Count (FC) for poisoning, asphyxiation, flammable/explosive gas, falling from height, and struck by objects. For these normally distributed datasets, the independent samples t-test (a parametric test) was utilized for analysis. The remaining datasets exhibited non-normal distributions and were analyzed using non-parametric methods in the subsequent steps. Prior to conducting the independent samples t-test, it is essential to verify the homogeneity of variance to ensure the statistical validity of the comparisons. Therefore, the Brown-Forsythe test was employed to assess the homogeneity of variance for the aforementioned eye-tracking metrics, with the null hypothesis positing equal variances across groups.

Table 8. Homogeneity test of homogeneity of variance.

Eye-Tracking Metrics	Brown-Forsythe (F)	p
TFF—Poisoning	0.266	0.61
TFF—Asphyxiation	0.486	0.492
FFD—Gas explosion	1.571	0.22
TFD—Poison	0.485	0.492
TFD—Gas explosion	0.004	0.952
TFD—Struck by objects	1.156	0.292
FC—Poisoning	0.004	0.949
FC—Asphyxiation	1.358	0.254
FC—Gas explosion	1.209	0.281
FC—Falling from height	2.376	0.134
FC—Struck by objects	0.83	0.37

presents the results of the homogeneity of variance test, including the test statistics and significance p-values.

The results of the homogeneity of variance test revealed that all relevant eye-tracking metrics, including Time to First Fixation (TFF) for poisoning and asphyxiation, met the assumption of equal variances (p > 0.05). Given the lack of statistical significance, the test failed to reject the null hypothesis. Consequently, the data satisfied the homogeneity of variance requirement, justifying the application of the independent samples t-test. To evaluate the significance of differences in eye-tracking metrics between the high- and low-risk propensity cohorts across various hazard types, statistical tests were conducted accordingly. The results for the metrics analyzed via the independent samples t-test are presented in

Table 9. Independent sample t test results of the influence of risk tendency on eye movement index.

Eye-Tracking Metrics	Group	Sample Size	Mean	SD	p
TFF—Poisoning	High-risk propensity	21	4.482	1.718	0.045
	Low-risk propensity	20	3.321	1.936
TFF—Asphyxiation	High-risk propensity	21	3.995	2.076	0.095
	Low-risk propensity	20	3.228	1.547
FFD—Gas explosion	High-risk propensity	21	0.252	0.139	0.433
	Low-risk propensity	20	0.287	0.099
TFD—Poisoning	High-risk propensity	21	1.458	0.818	0.516
	Low-risk propensity	20	1.659	0.853
TFD—Gas explosion	High-risk propensity	21	1.193	0.65	0.046
	Low-risk propensity	20	1.688	0.645
TFD—Struck by objects	High-risk propensity	21	1.964	1.416	0.932
	Low-risk propensity	20	2.004	1.133
FC—Poisoning	High-risk propensity	21	4.317	1.703	0.120
	Low-risk propensity	20	5.276	1.573
FC—Asphyxiation	High-risk propensity	21	3.28	1.323	0.757
	Low-risk propensity	20	3.443	1.511
FC—Gas explosion	High-risk propensity	21	4.443	1.628	0.032
	Low-risk propensity	20	5.603	1.184
FC—Falling from height	High-risk propensity	21	6.5	2.872	0.715
	Low-risk propensity	20	6.188	1.699
FC—Struck by objects	High-risk propensity	21	6.179	3.44	0.048
	Low-risk propensity	20	7.188	2.78

, while the results for those analyzed using the Kruskal-Wallis test are detailed in

Table 10. Kruskal-Wallis test results of the influence of risk tendency on eye movement indicators.

Eye-Tracking Metrics	Group	Sample Size	Mean	SD	p
TFF—Poisoning	High-risk propensity	21	4.665	3.43	0.262
	Low-risk propensity	20	3.841	0.942
TFF—Asphyxiation	High-risk propensity	21	4.977	3.701	0.019
	Low-risk propensity	20	3.01	1.845
FFD—Gas explosion	High-risk propensity	21	2.38	1.982	0.383
	Low-risk propensity	20	2.88	2.536
TFD—Poisoning	High-risk propensity	21	0.221	0.275	0.708
	Low-risk propensity	20	0.283	0.158
TFD—Gas explosion	High-risk propensity	21	0.367	0.3	0.708
	Low-risk propensity	20	0.369	0.229
TFD—Struck by objects	High-risk propensity	21	0.18	0.166	0.009
	Low-risk propensity	20	0.248	0.088
FC—Poisoning	High-risk propensity	21	0.19	0.377	0.967
	Low-risk propensity	20	0.19	0.234
FC—Asphyxiation	High-risk propensity	21	1.207	0.883	0.739
	Low-risk propensity	20	1.37	1.037
FC—Gas explosion	High-risk propensity	21	1.478	1.552	0.041
	Low-risk propensity	20	1.582	0.944

.

Based on the statistical analysis results presented in Table 9 and Table 10, the following conclusions can be drawn:

Significant differences were observed between the cohorts with different risk propensity traits regarding the Time to First Fixation (TFF) for poisoning and falling from height hazards (p = 0.045 < 0.05; p = 0.019 < 0.05). The low-risk propensity cohort detected poisoning and falling from height hazards within the scenes significantly faster.

Significant differences existed between the cohorts concerning the Total Fixation Duration (TFD) for flammable/explosive gas and falling from height hazards (p = 0.046 < 0.05; p = 0.041 < 0.05). Compared to the high-risk propensity cohort, the low-risk propensity cohort allocated more attention to these two types of hazards.

A significant difference was found between the cohorts in the First Fixation Duration (FFD) for the falling from height hazard (p = 0.009 < 0.05). For this hazard, the low-risk propensity cohort exhibited longer observation dwell times.

Significant differences were noted between the cohorts in the Fixation Count (FC) for flammable/explosive gas and struck by objects hazards (p = 0.032 < 0.05; p = 0.048 < 0.05). During the risk perception process, the low-risk propensity cohort demonstrated more active visual search behaviors, returning their attention to hazard information areas more frequently.

Figure 15 presents line charts of the gaze-related data for both high- and low-risk propensity cohorts across various hazard scenarios, comprising four subplots (a, b, c, and d) that illustrate participants’ gaze characteristics from the four dimensions of TFF, FFD, TFD, and FC, respectively. Overall, as shown in Figure 15a, inter-group comparisons reveal that the TFF of the low-risk propensity cohort across all hazard types was generally shorter than that of the high-risk propensity cohort, indicating that low-risk propensity individuals locate various hazards relatively faster. When comparing across hazard types, both cohorts exhibited relatively short TFFs for falling from height hazards; notably, the TFF of the high-risk propensity cohort for struck by objects hazards decreased drastically compared to other hazard types.

Figure 15b demonstrates that for the asphyxiation hazard, the FFD of both cohorts reached their respective peaks, with the low-risk propensity cohort scoring slightly higher than the high-risk propensity cohort. Only slight differences were observed for other hazard types; the low-risk propensity cohort showed relatively longer FFDs for poisoning, flammable/explosive gas, and struck by objects hazards, indicating a more concentrated initial focus on these risks. Figure 15c shows that both cohorts allocated a substantial amount of attention time to the struck by objects hazard. Figure 15d indicates that the low-risk propensity cohort attended to hazards more frequently, reflecting a higher degree of risk valuation through multiple visual inspections.

In summary, when confronted with hazards such as poisoning, asphyxiation, flammable/explosive gas, falling from height, and struck by objects, the high-risk propensity cohort exhibited longer TFF, relatively lower FFD, and lower TFD and FC. These findings collectively demonstrate that high-risk propensity participants possess a poorer risk perception capability, thereby supporting Hypothesis 1.

3.2. Incidence of Inattentional Blindness

Preliminary research has established that the high-risk propensity cohort exhibits weaker risk perception capabilities in potential hazard scenarios, specifically manifesting as abnormal eye-movement patterns (e.g., delayed first fixation, shorter fixation durations, and lower total fixation time). These phenomena directly point to deficits in attentional allocation. For instance, the high-risk cohort demonstrates a “delayed alertness–shallow processing” pattern; in the presence of concealed hazards (e.g., poisoning, explosive gas), their Time to First Fixation (TFF) is significantly longer than that of the low-risk cohort, yet their First Fixation Duration (FFD) is shorter. This delayed alertness may stem from attentional resources being consumed by the primary experimental task, leading to an unconscious neglect of potential risks.

Furthermore, both high- and low-risk cohorts exhibit instances of extremely short fixation durations or even a complete absence of fixation behavior toward certain hazard information areas. These observations suggest that Inattentional Blindness (IB) may be a critical variable leading to risk perception failure. Both cohorts may experience a “look-but-fail-to-see” phenomenon due to IB—even when the hazard source is foveated, insufficient cognitive resources are allocated for deep processing, resulting in an underestimation of the hazard’s probability. Consequently, this study selects individuals from Section 1 who exhibited short AOI fixation times or zero fixations to further explore the interrelationships among Inattentional Blindness, risk propensity, and risk perception capability.

In the actual context of confined space operations, the occurrence of inattentional blindness is significantly exacerbated by the harsh work environment. The specific microclimate within these workplaces—often characterized by extreme temperatures, high humidity, poor ventilation, and localized hypoxia—imposes a severe physiological and psychological burden on operators. This extreme microclimate accelerates cognitive fatigue and rapidly depletes working memory capacity, forcing the brain to narrow its attentional focus strictly to the primary task at hand, thereby drastically increasing the incidence rate of inattentional blindness toward peripheral safety hazards.

To investigate the significance of IB in the failure of safety risk perception, the proportion of IB-related failures was calculated. Analysis revealed that the average performance ratio for safety risk perception failure was 32.80%. According to the formula RPA = c/(a + b + c), the average proportion of participants noticing hazard information across all images was 59.40%. The proportion of IB within safety risk perception failures can be evaluated using the formula a/(a + b); this formula implies that perception failure occurs from the outset due to the non-detection of hazards. The calculated average proportion of IB across all hazards was 23.80%. This signifies that approximately 72.56% of the potential hazards that participants failed to perceive accurately were caused by inattentional blindness, indicating that participants failed to identify the risks from the very beginning while performing the primary task.

Specific results are illustrated in Figure 16. These findings demonstrate that IB plays a dominant role in the failure of safety risk perception. By calculating mean values, it was found that the probability of the high-risk propensity cohort experiencing IB was 2.55 times that of the low-risk propensity cohort. This substantial categorical difference is statistically supported by the highly significant positive correlation observed between continuous risk propensity scores and IB incidence (r = 0.625, p < 0.001), which will be further detailed in the subsequent correlation analysis (Section 3.3.1). This robust statistical evidence reveals a profound disparity in IB occurrence across different risk propensity traits.

3.3. Mediation Effect Analysis

3.3.1. Correlation Analysis

Prior to the analysis, normality tests were conducted for risk propensity, Inattentional Blindness (IB), and risk perception capability. The Shapiro-Wilk test results yielded significance values of p = 0.667, p = 0.336, and p = 0.171, respectively, indicating that all variables followed a normal distribution.

Utilizing SPSS 26.0, this study employed Pearson correlation analysis to examine the relationships among risk propensity, IB, and risk perception capability [22]. The absolute value of the Pearson correlation coefficient reflects the strength of the association between variables. The results of the correlation analysis between risk propensity traits, IB, and risk perception capability are presented in

Table 11. The correlation between risk tendency traits and inattentional blindness and risk perception ability.

	Risk Propensity	IB	Risk Perception Capability
Risk Propensity	1 (0.000)	-	-
IB	0.625 (0.000)	1 (0.000)	-
Risk Perception Capability	−0.403 (0.027)	−0.578 (0.001)	1(0.000)

.

Data in Table 11 indicate a significant positive correlation between risk propensity and Inattentional Blindness (IB) (r = 0.625, p < 0.01). Conversely, risk propensity and risk perception capability exhibit a negative correlation (r = −0.403, p < 0.05), and a significant negative correlation is also observed between IB and risk perception capability (r = −0.578, p < 0.01).

Figure 17 illustrates the pairwise relationships between risk propensity, IB, and risk perception capability. Figure 17a presents the scatter plot and regression line for risk propensity and risk perception capability; at a significance level of 0.027, it reflects a downward trend in risk perception capability as risk propensity increases. This finding is consistent with previous results, further demonstrating that the risk perception capability of the high-risk propensity cohort is significantly lower than that of the low-risk propensity cohort. Figure 17b depicts the relationship between risk propensity and IB, where the red regression line extends from the bottom-left to the top-right, indicating a positive correlation; that is, IB tends to increase as risk propensity levels rise. As shown in Figure 17c, the scatter plot reveals a statistically significant negative correlation between IB and risk perception capability. The regression line slopes from the top-left to the bottom-right, suggesting that a higher degree of IB is likely to reduce an individual’s risk perception capability. This reflects that when IB occurs, risk signals in the environment may be missed, leading to inaccurate or incomplete risk perception. Previous eye-tracking studies have shown that operators who exhibit fewer gaze returns and fewer fixations on hazard zones possess lower risk identification capabilities, a conclusion substantiated by the results in Figure 17c.

Cognitive and behavioral theories suggest that an individual’s risk propensity influences their information processing. Individuals with higher risk propensity may focus more on high-reward or stimulating elements in potential hazard scenarios while ignoring other critical information, thereby increasing the likelihood of IB. In turn, IB prevents individuals from acquiring risk information comprehensively and accurately, thus diminishing risk perception capability. Constructing a mediation model allows for the systematic integration of these theoretical perspectives, clearly presenting the internal logical connections between variables and helping to refine and expand existing theories. Analysis of the three scatter plots indicates that risk propensity is positively correlated with IB, which is negatively correlated with risk perception capability; risk propensity is also negatively correlated with risk perception capability. Based on these findings, a mediation model was established with risk propensity as the independent variable (X), IB as the mediator (M), and risk perception capability as the dependent variable (Y), aiming to explore the complex relationships among these factors. The construction of this model is based on the theoretical framework of the mediation effect path: “ X to M to Y, “ signifying that the independent variable (risk propensity) indirectly affects the dependent variable (risk perception capability) through the mediator (IB). The mediation model is illustrated in Figure 18.

This model aims to investigate the direct and indirect influence paths of risk propensity traits on risk perception capability, specifically testing the mediating effect of Inattentional Blindness (IB) in this relationship. Within this framework: Path A represents the first half of the mediation path, denoting the effect of risk propensity traits on IB; Path B represents the second half of the mediation path, denoting the effect of IB on risk perception capability; Path C represents the direct effect path of risk propensity traits on risk perception capability.

3.3.2. Mediation Model Testing

Given the subset sample size of 30 participants for the final risk perception assessment, a standard normal distribution of the indirect effect could not be strictly assumed. Therefore, to ensure robust statistical power and reliability against the small sample size, the mediating effect was tested using the bias-corrected bootstrapping resampling method (with 5000 resamples). Specifically, this approach evaluates whether the 95% confidence interval (CI) of the indirect effect product (Path a × b) includes zero. A significant mediating effect is established if the 95% CI does not contain zero; conversely, if the interval includes zero, it indicates the absence of a mediating effect.

While controlling for variables such as knowledge and work experience, the mediation analysis was conducted with risk propensity traits as the independent variable, inattentional blindness as the mediator, and risk perception capability as the dependent variable. The detailed results of the mediation test are presented in

Table 12. Mediation effect test results.

Path	Symbol	Effect Type	Estimate	Boot SE	p	95% CI
						Lower Limit	Upper Limit
A + B	a × b	Indirect effect	−0.123	0.165	0.006	−0.775	−0.127
A	a	X→M	0.152	0.034	<0.001	0.086	0.219
B	b	M→Y	−0.807	0.286	0.009	−1.368	−0.246
C	c′	Direct effect	−0.005	0.066	0.942	−0.134	0.124
C	c	Total effect	−0.128	0.056	0.030	−0.237	−0.019

.

The results indicated that the 95% confidence interval (CI) for the mediating effect ranged from −0.775 to −0.127, excluding zero. This demonstrates that inattentional blindness significantly mediates the effect of risk propensity traits on risk perception capability, with an indirect effect size of −0.123. Furthermore, the direct effect of risk propensity traits on risk perception capability was non-significant, whereas the indirect effect was significant. Specifically, upon introducing the mediating variable (inattentional blindness), the 95% CI for the direct path coefficient (c’) from X to Y included zero. In summary, the indirect influence of risk propensity on risk perception capability via inattentional blindness is validated, thereby supporting Hypothesis 2.

4. Discussion

Significant differences exist in risk perception capabilities among cohorts with different risk propensities. The low-risk propensity cohort exhibited a significantly shorter Time to First Fixation (TFF) on poisoning and falling-from-height hazards compared to the high-risk propensity cohort, indicating that the low-risk propensity cohort can detect these potential risks more rapidly. Regarding the Total Fixation Duration (TFD), the low-risk propensity cohort spent significantly more time fixating on flammable/explosive gas and falling-from-height hazards than the high-risk propensity cohort, demonstrating their heightened attention to these critical risks.

Furthermore, the study found that inattentional blindness (IB) plays a critical role in safety risk perception failures, accounting for approximately 72.56% of such failures. Correlation analyses revealed a positive correlation between risk propensity and IB, and negative correlations between risk perception capability and both risk propensity and IB. The mediation effect test demonstrated that risk propensity traits exert an indirect effect on risk perception capability primarily by influencing inattentional blindness. This exceptionally high proportion of perception failures attributed to IB significantly extends the findings of Park et al. [20], who observed IB in construction hazards. Our quantitative eye-tracking evidence confirms that in highly restrictive and complex confined spaces, individual risk traits significantly exacerbate this cognitive bottleneck, offering a novel perspective that differs from traditional hazard-exposure models. Additionally, the finding that individuals with a high-risk propensity tend to adopt an extensive but “superficial scanning strategy” provides a new dimension to existing literature on visual search behaviors. For instance, contrary to studies suggesting that experienced workers consistently employ efficient, goal-directed search strategies to locate hazards, our results reveal that intrinsic high-risk traits lead to a widespread but shallow distribution of attention, increasing vulnerability to cognitive omissions.

Stable personality traits significantly affect an individual’s ability to capture unexpected stimuli, which is consistent with the findings of Richards et al. However, previous research on inattentional blindness has predominantly relied on simple geometric tasks on computer screens (e.g., the traditional cross-tracking paradigm), and its external validity in complex engineering fields has been a subject of ongoing debate. By utilizing realistic confined space operation scenarios (e.g., poorly ventilated pipes and dark rooms with falling hazards) as experimental stimuli, this study further verified that in industrial environments with real physical characteristics, the inducing effect of risk propensity on inattentional blindness remains significant and frequently leads to the failure to detect critical risk information. This provides an explanatory dimension based on individual cognitive differences for understanding human error in complex human-machine systems.

The findings of this study offer valuable reference for safety management in modern high-risk industries, particularly in confined space operations. Traditional safety interventions predominantly focus on improving the physical environment, often overlooking the cognitive limitations of workers when processing information. First, clarifying the relationship between risk propensity and inattentional blindness helps safety managers recognize the innate differences among individuals in risk perception. Second, for the high-risk propensity cohort, conventional safety rule briefings may be insufficient to compensate for their attention allocation deficits in complex environments. It is recommended that future safety training incorporate simulation training or situational testing based on eye-tracking feedback to help workers identify their own visual search blind spots. This strategy aligns with the latest advancements in personalized safety training, which increasingly leverage eye-tracking and immersive virtual reality to enhance construction site hazard recognition [23,24]. This approach will guide them to rationally allocate cognitive resources while performing primary tasks, thereby effectively enhancing their capability to perceive hidden safety risks.

Limitations and Future Work

Despite its theoretical and practical contributions, this study has several limitations that warrant acknowledgment.

First, regarding the experimental sample and scope, the participants were university students majoring in safety engineering rather than professional confined-space workers. While this selection effectively controlled for confounding occupational variables (e.g., varying levels of prior safety training) and facilitated the baseline validation of the cognitive mechanism, this research should be strictly characterized as a mechanism-validation study rather than a direct operational safety guideline. Real-world operators may exhibit different perceptual adaptations due to chronic stress and extensive field experience.

Second, regarding the statistical analysis, the sample size (N = 41, with a subset of 30 for the final situation awareness test) is relatively small. However, it is important to clarify that our tested framework is a simple mediation model, and the subset of 30 was deliberately selected through strict oculomotor thresholds to ensure the analysis was based on high-fidelity cognitive responses. To rigorously address the statistical constraints of the small sample size, we employed bias-corrected bootstrapping with 5000 resamples to estimate the indirect effects, which provides robust confidence intervals without relying on normality assumptions.

Third, regarding practical application, while the proposed eye-tracking-based safety training solution can be integrated into regular onboarding programs in simulated environments, widespread deployment faces practical constraints. These include the high procurement costs of wearable equipment, the reliance on specialized personnel for data interpretation, and the inherent difficulty of perfectly replicating the extreme psychological stress and harsh microclimate of a real confined space within a safe training simulator.

Therefore, future research should aim to conduct large-sample cohort studies replicating this experimental paradigm with frontline workers in highly ecologically valid operational settings, thereby verifying the generalizability of the proposed mediation model and exploring more cost-effective methods for safety interventions.

5. Conclusions

This study successfully integrated risk propensity, inattentional blindness, and risk perception into a unified cognitive evaluation framework within the context of confined space operations. Utilizing eye-tracking technology, we objectively quantified the visual search discrepancies among operators with different risk traits. The findings reveal that individuals with high risk propensity predominantly adopt a “delayed alertness–shallow processing” strategy, characterized by significantly longer Time to First Fixation and shorter fixation durations on critical hazards. Crucially, the mediation analysis verified that inattentional blindness accounts for approximately 72.56% of safety perception failures and serves as a full mediator between risk propensity and risk perception capability. These results highlight that cognitive resource allocation deficits, driven by inherent risk traits, are primary precursors to human error in complex environments. This research provides a robust theoretical foundation for shifting industrial safety management from traditional behavioral compliance to proactive, cognitive-based interventions, suggesting that eye-tracking simulation training could substantially enhance operators’ hazard detection efficiency and overall occupational sustainability.