Automated Assessment of Construction Workers’ Accident Risk During Walks for Safety Planning Based on Empirical Data

Abstract

Ensuring workers’ safety is a critical component of social sustainability in the construction industry. Accidents that occur while workers are walking on construction sites constitute a significant portion of overall accidents, yet they are often overlooked in conventional task-oriented safety risk assessments. This study proposes novel Accident-During-Walk (ADW) risk indices, hierarchical and data-driven metrics designed to quantify workers’ accident risk during walks. The indices are built on Association Rule Mining and utilize structured accident data, accounting for both environmental and work-related attributes. By integrating these indices with project-specific work schedules and worker allocation plans, this study establishes an automated method for daily and weekly look-ahead ADW risk monitoring aligned with construction progress. Case studies on two construction projects validate the discriminative power of the proposed method. The results demonstrate that the indices effectively capture risk fluctuations driven by concurrent multi-trade operations and environmental severity. Notably, the analysis reveals counterintuitive patterns where adverse weather conditions paradoxically reduce risk values by constraining worker mobility, a nuance often missed by static assessments. Ultimately, this framework serves as a data-driven decision-support tool, enabling safety managers to transition from uniform inspections to targeted interventions during high-risk periods, thereby fostering a safer and more socially sustainable construction environment.

1. Introduction

Achieving sustainable construction requires balancing environmental, economic, and social responsibilities. Among these, ensuring occupational health and safety is a cornerstone of social sustainability. However, fulfilling this social obligation remains difficult due to the industry’s inherently hazardous nature. The construction industry is recognized as a high-risk sector, accounting for 20.3% of all industrial accidents in the United States as of 2023 [1]. This persistent risk has motivated extensive research into accident causality and the establishment of organized database systems [2]. In response, multiple countries have established structured accident reporting systems through the implementation of Reporting of Injuries, Diseases, and Dangerous Occurrences Regulations (RIDDOR) and the Occupational Injury and Illness Classification System (OIICS). These accumulated databases serve as vital resources for root cause analysis and prevention strategies [3,4]. Furthermore, structured accident data supports safety management practices, including Job Hazard Analysis (JHA), which utilizes the work breakdown structure (WBS) of construction projects to identify task-specific occupational risks [5].

Despite these advancements, traditional risk assessments often overlook hazards associated with transient activities. Construction sites are dynamic environments characterized by frequent worker movements—such as transitions between tasks, breaks, or shifts in location—that differ significantly from stationary work. Studies indicate that accidents during these movements constitute a substantial portion of site incidents. Ayhan and Tokdemir [6] revealed that 31% of construction accidents occurred while workers were walking, often due to a lack of hazard awareness. Similarly, Yi [7] noted that approximately 20% of accidents occur during walking rather than active tasks.

These Accidents During Walks (ADW) refer to accidents that occur during off-task movement of workers, such as transitions at the start, during breaks, and at the end of work. The types of events associated with ADW encompass not only slips, trips, and falls (STFs) but also collisions with objects. Given that STFs account for 39% of fatal injuries in construction [1,8], distinguishing ADW from Accidents During Tasks (ADT) and developing specialized management approaches is imperative [9]. Currently, safety managers rely heavily on intuition and experience to assess ADW risks, considering factors such as climate, terrain, and workforce allocation. While measures like securing passageways and improving lighting are implemented, managers face significant challenges in developing quantitative daily or weekly safety plans.

Such challenges stem from two main factors. Firstly, there are limited indices that facilitate the provision of ADW risk information during the project execution. Secondly, there is a lack of automated methods to quantitatively assess the risk level of ADW using accumulated accident data. Structured risk assessment methods during project execution, such as JHA, are based on task-specific hazard identification derived from the WBS of construction projects [5]. These traditional hazard identification practices rely on the judgment of experienced professionals, who examine the activities and conditions of the work environment [10]. However, reliance on qualitative judgements is limited in identifying accident hazards associated with transient worker movements between tasks or outside designated work areas. Consequently, despite the availability of extensive accident databases, there is a critical lack of automated methodologies that reorganize this empirical data for ADW risk assessment.

To address this gap, this study proposes an automated method that provides safety managers with risk levels for ADW and hazardous objects for each time interval (e.g., weekly or daily) based on the WBS of the project schedule. This study is grounded in the hypothesis that ADW risks stem from identifiable combinations of environmental conditions and workforce allocation, which can be quantified using data-driven metrics. To enable schedule-based delivery of ADW hazard information, this study reexamines the metrics commonly employed in association rule mining (ARM) algorithms. This rule-based machine learning approach identifies item sets that co-occur in large-scale databases [11]. The approach infers association rules based on the likelihood of item co-occurrence, as quantified by metrics such as support, confidence, and lift. While data-driven ARM approaches have been consistently adopted to analyze accident influence factors [12,13,14], they have neither sufficiently addressed ADW nor provided quantitative evaluations of its risk levels. The proposed ADW risk indices, are therefore a set of hierarchical, quantitative indices that use ARM metrics, allowing their integration into project schedules.

To support the calculation of these indices, this study presents a corresponding database structure and a reorganization process for the accident case database to incorporate empirical project data such as weather forecasts and workforce allocation plans. The effectiveness of the proposed approach is validated through two case studies, specifically by evaluating whether the derived indices accurately reflect risk fluctuations due to seasonal weather and trade overlaps. Additionally, the proposed method allows for the calculation of continuously updated results as new accident data are collected. Ultimately, by enabling proactive look-ahead safety planning, this study aims to enhance construction site safety and generate economic benefits by preventing costly project delays and compensation claims associated with accidents.

2. Literature Review

2.1. Breakdown Structure of Construction Accident Databases

Hazard identification and risk assessment are fundamental to safety management. Their effectiveness depends on accurate causal models of accidents, which in turn require reliable reporting systems [15]. However, accident data collected by construction companies is often poorly reported, hindering integration and analysis [16]. To improve consistency, national initiatives such as OIICS and RIDDOR have standardized reporting since the 1990s [15,17]. For example, OIICS 3.02 classifies accidents into five categories: nature of injury or illness, part of body affected, event or exposure, source of injury or illness, and worker activity and location [18].

Nevertheless, these systems inadequately reflect construction-specific traits such as frequent crew changes, weather impacts, and reliance on temporary labor [5]. These dynamic factors create a significant gap between standardized reporting codes and the complex reality of daily site operations, often leading to the oversimplification of accident contexts. To address these gaps, numerous studies have examined causal processes in greater detail [3,6]. For instance, Ayhan and Tokdemir [6] applied clustering and neural networks (i.e., Latent Class Clustering and Artificial Neural Networks, LCCA and ANN), finding that approximately 31% of cases were ADW, primarily occurring when workers failed to recognize hazards in their walking environment. This highlights the limitations of standardized databases in providing sufficient information to evaluate ADW risk among construction workers.

For example, the Census of Fatal Occupational Injuries (CFOI) categorizes accident types under the “event or exposure” domain, where STFs are identified as the primary accident modes [19]. However, because ADW refers to accidents during non-task walking, it also encompasses various contact accidents. Moreover, while the “worker activity and location” category includes codes related to movement, many involve interactions with materials or equipment, making it challenging to distinguish non-task walks from other activities. Overall, current classification structures are insufficient to capture ADW hazards, underscoring the need for alternative analytical methods.

2.2. Safety Risk Assessment in Construction

Given the dynamic and complex nature of the construction industry, many studies have explored specialized approaches to safety risk assessment. Hide et al. [3] analyzed 100 unstructured accident reports to identify a hierarchical structure of causal influences, including management, site environment, and individual factors. Their work emphasized qualitative elements such as worker attitudes, weather, spatial layout, and material conditions. These findings reinforced the view that construction safety risk assessment should integrate diverse qualitative factors and translate expert judgments into quantitative measures.

For example, Pinto [10] developed the Qualitative Occupational Risk Assessment Model (QRAM), which quantifies expert evaluations of major accident types using fuzzy set theory. Jannadi and Almishari [20] introduced an approach that incorporates accident databases into risk evaluation, defining scores as the product of probability, severity, and exposure. Accident frequencies were used as proxies for probability, severity was inferred from compensation costs, and exposure was adjusted through expert judgment. More advanced methods include the Construction Hazard Assessment with Spatial and Temporal Exposure (CHASTE) [5,21]. This framework evaluates risks by detecting overlaps between source and victim activities, focusing on “loss-of-control” events derived from structured databases and digital construction plans such as Building Information Models (BIM). This method requires detailed activity-level data and provides more spatially precise assessments [22,23]. Collectively, these studies suggest that activity-based analyses are the dominant approach in construction safety risk assessment. As a result, risks associated with non-working states, such as ADW, are often overlooked. Additionally, advanced methods that incorporate spatial analysis require custom databases or modifications of existing ones to ensure compatibility.

Because construction accident databases primarily contain categorical attributes [15,17], ARM has become a common tool for identifying complex causal relationships in construction safety research [12], for example, used ARM on a national accident database in Taiwan to reveal frequent scenarios involving multiple factors, such as unsafe contact, small company size, and lack of protective equipment that lead to falls. Similarly, Korean studies have applied ARM to find patterns in fatal accidents at small-scale construction sites [24,25]. Yao et al. [14] advanced this approach by integrating ARM with Bayesian Networks (BNs) to construct a risk assessment model that identifies probable accident paths and sensitive factors. To strengthen the credibility of ARM-derived scenarios, Huo et al. [13] validated rule-based results against expert evaluations and assessed their similarity. While these studies effectively employed ARM thresholds to extract causal patterns, they did not utilize the metrics as quantitative measures of accident risk. Consequently, the application of ARM has largely remained retrospective, lacking the necessary framework to translate these patterns into predictive risk indices for daily safety management.

To systematically identify the limitations of the aforementioned approaches from the perspective of ADW risk assessment,

Table 1. Comparison of existing risk assessment methods in the context of ADW.

Category	Specific Methods	Key Characteristics	Limitations
Qualitative risk assessment	JHA [5], QRAM [10]	- Relies on expert experience and subjective judgment. - Focuses on task-specific hazards.	- Dependent on individual safety managers’ experience. - Overlooks ‘transient’ risks such as ADW.
Quantitative risk assessment	Data-driven risk matrix [20]	- Calculates risk as probability × severity. - Uses historical frequency as a proxy.	- Uses average probabilities, ignoring daily variations in project-specific conditions.
Spatial-Temporal Analysis	CHASTE [21]	- Analyzes spatial conflicts and overlaps, focusing on collisions.	- Requires heavy data processing - Requires detailed schedule and space models.
	BIM-based risk analysis [22]	- Visualizes hazards in 3D/4D using BIM	- Primarily target falls, not worker slips/trips.
Hybrid machine learning	LCCA and ANN [6]	- Predicts accidents based on complex patterns	(ANN) - Prediction logic remains opaque. - Focuses on severity classification rather than occurrence probability.
	ARM-based BNs [14]	- Handles non-linear relationships using hybrid approaches.	(Bayesian networks) - Constructing network structure and prior probabilities is complex - Difficult to update dynamically.
Rule-based pattern mining	ARM [12,24,25]	- Extracts explicit if-then rules.	- Not integrated with daily work schedules for proactive planning.
	ARM-based critical path analysis [13]	- Identifies frequent accident scenarios and causal paths. - Validates rules against expert opinions.	- Focuses on identifying static causal mechanisms rather than quantifying daily risk levels.

presents a comparison of representative risk assessment methodologies. As summarized in the table, traditional qualitative and quantitative methods often lack the context sensitivity needed to capture transient risks such as ADW. Additionally, while advanced data-driven models provide predictive capabilities, they frequently face challenges related to interpretability or a lack of quantitative integration with daily project schedules. Therefore, the objective of the proposed method in this study is to bridge these gaps by introducing a hierarchical, schedule-integrated approach based on ARM.

3. Indices of Construction Workers’ Accident Risk During Walks

3.1. Definition of Hierarchical ADW Risk Indices

ARM, first introduced by Agrawal et al. [11], is a rule-based machine learning technique developed initially to identify frequent item sets in retail databases. Using the Apriori algorithm [26], ARM derives association rules based on metrics such as support and confidence, later enhanced by lift [27] and other measures [28,29].

In assessing ADW risk, ARM offers distinct advantages over predictive or probabilistic models. Predictive methods, such as ANN, primarily focus on classifying outcomes based on non-linear patterns. However, these methods often function as ‘black boxes,’ lacking the interpretability required to identify specific combinatorial hazards [6]. Similarly, while BNs are effective for causal inference, they require complex prior probability estimation and fixed network structures, making them difficult to integrate dynamically with changing work schedules [14]. In contrast, ARM is an unsupervised method that efficiently identifies interpretable item sets without enforcing strict structural assumptions or requiring predefined targets.

In ARM, support measures the probability that both an antecedent itemset (X) and a consequent itemset (Y) appear together in the transaction database (Equation (1)). Either itemset may contain multiple items. When one of the item sets is empty, support reduces to the probability of the remaining itemset alone (e.g., $Support\left(X\right)=P\left(X\right)$). Confidence refers to the conditional probability of Y occurring given X, as presented in Equation (2). Lift indicates the increased likelihood of Y occurring when X is present compared to when X is absent. It compares the observed co-occurrence of X and Y with the expected probability under independence across the entire dataset. A lift greater than 1 signifies a positive association between X and Y. For example, a lift of 1.25 means that the joint occurrence of X and Y is 25% more probable than if they were independent. A notable property of this metric, as expressed in Equation (3), is its invariance to the order of the antecedent and consequent [14].

$$Support\left(X\Rightarrow Y\right)=P(X\cap Y)$$

(1)

$$\mathit{Confidence}\left(X\Rightarrow Y\right)=P\left(Y∣X\right)={\displaystyle \frac{Support(X\cap Y)}{Support\left(X\right)}}$$

(2)

$$Lift\left(X\Rightarrow Y\right)={\displaystyle \frac{P\left(Y∣X\right)}{P\left(Y\right)}}={\displaystyle \frac{Confidence\left(X\Rightarrow Y\right)}{Support\left(Y\right)}}={\displaystyle \frac{Support(X\cap Y)}{Support\left(X\right)\times Support\left(Y\right)}}$$

(3)

This study assumes a database in which each row represents a construction accident case, and the columns correspond to categorical attributes describing the accident conditions, as shown in Figure 1. The database necessarily includes an attribute indicating whether the accident occurred during transitions/movements or during tasks (i.e., ADW/ADT status). Other attributes are classified into ‘scheduling-related conditions (SC)’, which capture WBS-based scheduling information, and ‘environmental conditions (EC)’, which cover all remaining conditions.

From an ARM perspective, each accident corresponds to a transaction, and item sets represent combinations of potential causal factors. In this context, the lift value serves as a quantitative indicator of how accident probability changes when conditions coincide, relative to their independent occurrences.

The proposed indices are termed “hierarchical ADW risk indices” because they are derived by sequentially forming item sets across attributes (e.g., pairing an item from EC #1 with one from EC #2, and so on). To define ADW risk indices, the first attribute is confined to the ADW/ADT status (#0). An itemset is then constructed by pairing one item from EC #1 with ADW, and its lift value is defined as the Level 1 Environmental Condition Risk (${ECR}_{L1}$) within the ADW risk index hierarchy. In other words, ${ECR}_{L1}$ can be expressed in Equation (4).

$${ECR}_{L1}=Lift\left(Item in EC \#1\Rightarrow ADW\right)$$

(4)

Next, a filtered dataset is constructed that consists exclusively of cases classified as ADW. Within this subset, the lift value between an item in EC #2 and an item in EC #1 can be calculated. This metric quantifies the increased likelihood of these two conditions occurring together in ADW cases relative to the assumption of independence. Consequently, the Level 2 ECR (${ECR}_{L2}$) is defined as follows:

$${ECR}_{L2}={ECR}_{L1}\times Lift\left(Item in EC \#2\Rightarrow Item in EC \#1\right)$$

(5)

For example, suppose EC #1 and EC #2 denote weather conditions and temperature levels, respectively. If ${ECR}_{L1}\left(Rainy\right)$ is 1.5, indicating that the likelihood of an ADW occurrence under rainy conditions is 50% higher compared to independent occurrences. The lift value of ‘Cold’ given ‘Rainy’ is 1.20, then an ${ECR}_{L2}={ECR}_{L1}\left(Rainy) \times Lift(Cold\Rightarrow Rainy\right)=1.8$ signifies that the likelihood of an ADW occurrence under cold and rainy conditions is elevated by 80% compared to other circumstances.

By iterating this process, the lowest-level ${ECR}_{Lx}$, which fully incorporates considerations for $x$ environmental conditions, enables safety managers to primarily assess the ADW risk index resulting from the combination of environmental factors, independently of worker allocation based on the work schedule.

3.2. Incorporating Worker Allocation into the ADW Risk Indices

${ECR}_{Lx}$ quantify the impact of combined environmental factors on accident occurrence. However, they are not directly associated with work characteristics that change over the course of the project. During project execution, worker allocation plans are established based on the project’s WBS. When SC attributes include details of specific works or activities in which workers participated, these can be integrated to derive project-specific risk indicators.

At this level, the ADW risk indices are defined as the Scheduling-related Condition Risk (${SCR}_{Lx}$). Similar to ${ECR}_{Lx}$, ${SCR}_{Lx}$ are constructed stepwise, but they incorporate worker allocation through an accumulated weighted sum. Specifically, for each work or activity $i$, the number of workers assigned ($W_i$) is multiplied by the corresponding accident risk (${SCR}_{Lx,i}$), and the products are summed to obtain the total accident risk volume. Accordingly, ${SCR}_{Lx}$ is defined as follows:

$${SCR}_{Lx}=\sum _{i=1}^n(W_i\times {SCR}_{Lx, i})$$

(6)

where $n$ represents the total number of works considered. This approach assesses the overall risk volume to which the workforce is exposed, reflecting the distribution of workers across activities.

Nonetheless, the accumulated weighted sum method has a limitation: the outcomes are significantly affected by variations in project size and workforce allocation, rendering it unsuitable for unbiased comparisons of risk levels across projects or for assessing the average risk per worker. Particularly, tasks with larger workforces may disproportionately distort the overall risk evaluation.

Therefore, in this study, where worker allocation planning exists at certain levels, the ${SCR}_{Lx}$ is redefined through the application of a weighted average. Equation (7) normalizes the aggregate weighted sum by the total number of workers, thereby facilitating a fair comparison of the average accident risk levels across various construction projects, regardless of differences in workforce size.

$${N-SCR}_{Lx}={\displaystyle \frac{{\sum }_{i=1}^n(W_i\times {SCR}_{Lx,i})}{{\sum }_{i=1}^n{W}_i}}$$

(7)

4. Assessment of ADW Risk Based on Empirical Data

4.1. Automated ADW Risk Assessment Process

This section outlines the procedure for calculating and presenting project-specific ADW risk information based on empirical data. Figure 2 illustrates the interaction process between the safety manager and the proposed automated ADW risk assessment system, depicted using a Unified Modeling Language activity diagram.

The process can comprise the following four steps: (1) Hierarchical Lift (HL) database construction, (2) Risk-condition querying, (3) Risk index computation, and (4) Risk reporting & visualization.

▪

HL database construction: the first stage involves reshaping the reference accident case database to construct a new database that contains lift value in a hierarchical branching-tree structure (hereafter referred to as Hierarchical Lift, HL). At this stage, the safety manager must prepare a database with a structure similar to that shown in Figure 1, which will serve as a reference for assessing ADW risk in the project. The database is reduced to retain only the attributes that the safety manager intends to consider. In this hierarchical arrangement, ECs are positioned at higher levels, while SCs are positioned at lower levels. This ensures that the computation of $ECRs$ precedes that of $SCRs$. During the reshaping process, the items of each attribute are reorganized into a branching tree structure, and the frequency of occurrence is recorded at each branch. These reshaped frequency data are subsequently used to compute HL. Figure 3 presents the conceptual schema of the database after reshaping and HL computation are complete. The HL database is represented as a branching tree composed of items from each attribute. Subtotals under given conditions are first calculated, after which the ARM metrics of item pairs between antecedent and consequent attributes are computed. The lift values of each branch become the primary targets for querying in subsequent steps; however, other ARM metrics, as shown in Figure 3, are also computed as prerequisite data. For example, when calculating the lift of branch ‘Item A’, the data within scope (i) are used, whereas for branch ‘Item a’, the relevant data are drawn from scope (ii).

▪

Risk-condition querying: the second stage involves feeding weather forecast data into the system and querying the conditions that align with the project schedule. In this step, the safety manager defines the time window for risk analysis and inputs the corresponding work schedule into the system. Based on the provided schedule and the retrieved environmental data, the system queries the HL database to extract the HL values that match the specified conditions. Once the project work schedule is established, the project manager formulates a worker allocation plan for each scheduled task, which will serve as the basis for index computation in the subsequent stage.

▪

Risk index computation: the third stage focuses on calculating the ADW risk indices for the planned work activities within the defined analysis time window. For this computation, the safety manager must provide the system with information on worker allocation for each WBS element of the work schedule. Using the risk-condition pairs obtained in the previous step together with the worker allocation data, the indices defined in Equations (4)–(7) are computed for each time unit (e.g., day, week). In parallel, this stage concurrently identifies the list of ADW-inducing objects.During the preparation of the reference database, specific attributes, such as accident-inducing objects, may contain a large number of items. If such attributes are positioned at intermediate levels of the branching tree, they can generate an excessive number of branches, many of which have no associated accident cases. This leads to undesirable fragmentation of the tree structure. To mitigate this issue, these attributes are positioned at the lowest level of the branching tree and excluded from the risk index computation. Instead, a list is generated for items with an accident frequency of at least one, ensuring that they are still captured for interpretive purposes without compromising the efficiency of index calculation. Further details of this stage can be more clearly understood through the example system interface, which will be presented later.

▪

Risk reporting & visualization: the final stage involves aggregating the computation results of the ADW risk indices and delivering them to the safety manager in a comprehensible form. Since the third step computes the indices repeatedly for each time unit, this stage aggregates the outputs across iterations and organizes them into an integrated report. Time series visualizations are particularly compelling for conveying these results, as they allow managers to observe changes in risk levels over the project timeline. Examples of such visual outputs will be presented in the subsequent case studies.

4.2. Example Interface of the ADW Risk Assessment System

As the input advances through each stage, a cumulative ADW risk level is calculated and depicted as a bar chart at the top of Figure 4. Furthermore, based on the input conditions, the interface identifies and presents a list of hazardous objects that have been historically associated with ADW. This feature helps safety managers identify key risk factors under similar conditions.

The example interface demonstrates the ADW risk level under weather conditions classified as ‘Mild’ temperatures and ‘Rainy’ weather. Under these specific circumstances, the interface indicates that ADW has historically occurred 55 percent more frequently than under other environmental conditions, as represented by the baseline level depicted as a red dashed line. Furthermore, when workers are assigned to the activities shown in Figure 4 under these conditions, the cumulative ADW risk level exceeds four times the baseline. This increased risk is explicitly reflected in the sequential bar chart. This interactive interface exemplifies the capacity of the proposed method to transform structured accident data and project-specific input into interpretable indicators of ADW risk.

5. Case Studies

5.1. Description of the Applied Database and Case Projects

For the construction of the HL database, this study utilized 23,468 accident cases accumulated over five years (2019–2024) from the Construction Safety Information System (CSI) database [30]. Since the raw data lacked labels indicating ADW status, a text-mining-based classification method was employed to categorize cases as either ADW or ADT [31]. This supervised learning model, using Bidirectional Encoder Representations from Transformers (BERT) embeddings and a Naive Bayes classifier, was validated on a manually labeled dataset of 1360 ground-truth instances, achieving a precision of 93.7% for the ADW class. Consequently, 7720 ADW and 10,582 ADT cases served as the foundational data for the HL database.

This process demonstrates a hybrid approach that leverages the strengths of different data-driven techniques. While pre-trained language models like BERT excel at semantic interpretation and rapid hazard identification from unstructured text [32,33], they often present challenges in providing consistent quantitative metrics due to their probabilistic nature. In contrast, for the subsequent phase of integrating risk with rigid construction schedules, deterministic reproducibility is essential. Therefore, this study uses natural language processing to structure the data and transitions to the ARM-based approach to compute mathematically consistent risk indices.

Following data collection, attributes for the HL database were selected based on theoretical relevance and data availability. Previous studies have highlighted that construction safety risks are significantly influenced by environmental factors, such as weather conditions [3], and work-related factors linked to the WBS [5]. Accordingly, attributes available in the CSI database that could be structured as categorical variables to meet ARM requirements were selected and preprocessed as summarized in

Table 2. Classification of attributes and items applied in case studies.

HL Database Attribute	Case Study Application
	Attribute	Item
EC #1	Weather condition	Clear; Rainy; Snowy; Foggy; Windy
EC #2	Temperature level	Extreme cold (below 0 °C); Cold (0–10 °C); Mild (10–26 °C); Heat (26–33 °C); Extreme heat (above 33 °C)
SC #1	Work type	Foundation and pavement work; Interior work; Metal, window, door, and roofing work; Painting, waterproofing, masonry and stone work; Landscaping and miscellaneous facility work; Concrete and rebar work; Demolition and scaffolding work; Water supply work; Railway work; Steel structure work; Dredging work; Mechanical work; Piping and gas supply work; Electrical work; Incidental work
SC #2	Activity type	Traveling; Excavation and piling; Equipment handling; Tamping and grading; Painting and plastering; Plumbing and wiring; Cladding and laying; Assembling and installing; Welding; Transporting and lifting; Preparation and inspection; Cleaning and tidying; Pouring and curing; Demolishing; Miscellaneous

. Numerical temperature data were converted into categorical items based on regional legal standards and reference points for Wind Chill and Heat Indices [34,35]. Similarly, the work and activity types were reorganized into 15 standardized items each to align with regional construction industry classifications [36].

Table 3. Summary of case study project characteristics.

Project Details	Case A	Case B
Project type	Educational facility	Office building
Construction period	26 months	29.8 months
Floor configuration	B1–6F	B4–15F
Site area (m2)	19,866	5713
Building area (m2)	4173	3340
Gross floor area (m2)	19,723	43,410
Building coverage ratio	0.21	0.58
Floor area ratio	0.80	4.61

summarizes the key characteristics of the selected projects for the case studies. Two projects with distinct site conditions were chosen for validation: Case A (an educational facility) and Case B (an office building). As detailed in the table, Case B has a significantly larger gross floor area (43,410 m²) than Case A (19,723 m²), resulting in a higher workforce density despite similar construction durations.

The ADW risk assessment focused on five working days at every 10% progress interval (10–90%). Figure 5 and Figure 6 present the project-specific input data, capturing seasonal weather variations over the two-year periods and daily workforce allocation.

Notably, while early stages (10–40%) involved minimal trade overlap, later stages experienced significant densification due to concurrent multi-trade activities. Case B reached a peak workforce of approximately 250 personnel during these overlaps, whereas Case A peaked at around 100, highlighting the substantial difference in site congestion levels.

5.2. Results of ADW Risk Indices Calculation

To validate the methodology, the time series variations of three hierarchical indices were analyzed: ${ECR}_{L2}$, ${SCR}_{L1}$, and ${SCR}_{L2}$. The assessment commences from ${ECR}_{L2}$ to explicitly incorporate the temperature effects alongside weather conditions. ${SCR}_{L1}$ signifies the weighted ADW risk considering workforce allocation at the work level, whereas ${SCR}_{L2}$ reflects the weighted risk at a more detailed activity level. Figure 7 and Figure 8 illustrate these indices—presenting both unnormalized and normalized values—derived from the input data. The key findings from the comparative analysis are summarized in

Table 4. Key findings from ADW risk index calculations.

Findings	Supporting Evidence
	Description	Project	Time Unit		Index
Capability to assess ADW risk through work schedule integration	Increased values at lower levels of the ARW indices	Case A	40%	Day 3	${ECR}_{L2}$ $N-{SCR}_{L1}$ $N-{SCR}_{L2}$	1.55 1.77 4.46
	Decreased values at lower levels of the ARW indices	Case A	90%	Day 1	${ECR}_{L2}$ ${N-SCR}_{L1}$ ${N-SCR}_{L2}$	1.00 0.90 0.97
Elevated ADW risk associated with trade overlap	Sharp increase in ${N-SCR}_{L2}$ resulting from finishing-work overlap	Case A	30%	Days 1 to 5	${N-SCR}_{L2}$	1.54 (average)
		Case A	40%	Days 1 to 5	${N-SCR}_{L2}$	2.52 (average)
Identification of activities with higher ADW risk under clear and mild conditions	Higher ${N-SCR}_{L2}$ on ‘Clear’ and ‘Mild’ days during foundation work	Case A	10%	Day 2	${ECR}_{L2}$ ${N-SCR}_{L2}$	0.99 5.72
	Lower ${N-SCR}_{L2}$ on ‘Clear’ and ‘Heat’ days during foundation work	Case A	10%	Day 1	${ECR}_{L2}$ ${N-SCR}_{L2}$	1.00 2.95
	Lower ${ECR}_{L2}$ and ${N-SCR}_{L2}$ for interior works under ‘Rainy’ and ‘Cold’ conditions	Case A	70%	Day 1	${ECR}_{L2}$ $N-{SCR}_{L2}$	0.43 0.43
Extreme ADW risk in foundation works under harsh conditions	Higher ${N-SCR}_{L2}$ of foundation works on ‘Snowy’ and ‘Extremely cold’ days	Case B	10%	Day 3	${ECR}_{L2}$ $N-{SCR}_{L2}$	15.04 27.34
	Lower ${N-SCR}_{L2}$ of concrete works on ‘Snowy’ and ‘Extremely cold’ days	Case B	50%	Day 5	${ECR}_{L2}$ ${N-SCR}_{L2}$	15.04 14.62
Capability to detect high-risk activities within trade overlaps	${N-SCR}_{L2}$ Changes due to the combination of rainfall with structural and interior works	Case B	60%	Day 2	${ECR}_{L2}$ ${N-SCR}_{L2}$	1.00 1.68
		Case B	60%	Day 3	${ECR}_{L2}$ ${N-SCR}_{L2}$	1.55 4.15

.

A notable transition in ${N-SCR}_{L2}$ is observed in Case A (Figure 7) between the 30% and 40% intervals. While it remained stable during structural work, a sharp increase was observed at the 40% progress as finishing work began concurrently. This pattern, less evident in ${N-SCR}_{L1}$, becomes clearer when combined with detailed activity information, enabling safety managers better to recognize elevated ADW risk from concurrent multi-trade operations.

During the 10% to 20% interval, significant variability in ${N-SCR}_{L2}$ was driven mainly by weather conditions rather than workforce levels. Lower ADW risk values appeared on ‘Heat’ and ‘Extreme cold’ days, while higher values were recorded under ‘Mild’ conditions. This indicates that adverse weather restricts workers’ mobility, thereby reducing risk and the likelihood of accidents. A similar trend is also evident on Day 1 of the 70% interval, where extremely cold and rainy conditions during interior works resulted in the lowest values (0.43). These counterintuitive findings underscore the importance of data-driven assessments in interpreting weather-activity interactions.

In Case B (Figure 8), ${SCR}_{Lx}$ values are more heavily influenced by workforce scale, indicating that higher workforce density escalates ADW risk through site congestion. However, environmental severity can outweigh workforce scale. The most prominent outliers (i.e., Days 2 and 3 at the 10% interval, and Day 5 at the 50% interval) occurred under snowfall and extreme cold. Despite only six workers performing foundation work, risk levels were extreme, whereas finishing work at the 50% interval was curtailed under similar conditions, preventing escalation. These findings emphasize the need for conservative thresholds when managing external works, such as foundation construction.

The analysis of the 60% interval further highlights the method’s ability to detect activity-level variations within overlapping trades. While ${N-SCR}_{L2}$ remained below 2 on Days 1 and 2, rainfall on subsequent days raised values to approximately 3.4. The peak value (4.2) on Day 3 was associated with painting, plastering, and cleaning—activities involving frequent movement and material handling. These results demonstrate that the proposed indices not only capture the amplifying effect of rainfall under multi-trade conditions but also pinpoint specific activities that pose heightened ADW risks.

To statistically validate the discriminative power of the proposed indices, the distribution of index values was analyzed using valid data points from the 45-day schedule, excluding non-working days.

Figure 9 presents the resulting distribution and descriptive statistics for both cases. The results show that each level of the index hierarchy has a specific role in risk detection. ${ECR}_{L2}$ remained clustered near the baseline (1.0), validating its function as a fundamental indicator of weather-driven risk fluctuations. In contrast, integrating workforce allocation (${N-SCR}_{L1}$) and activity details (${N-SCR}_{L2}$) significantly widened the risk distribution. In Case B, ${N-SCR}_{L2}$ showed a substantial interquartile range (IQR) of 0.95, nearly 47 times broader than that of ${ECR}_{L2}$. This expanded variance confirms that combining environmental data with specific activity details captures complex interaction effects, offering a more discriminative risk profile than environmental assessment alone.

Building on the statistical evidence of index sensitivity, this study proposes a practical application framework with different thresholds for each risk index. For macroscopic indices (i.e., ${ECR}_{L2}$ and ${N-SCR}_{L1}$), a baseline threshold of 1.0 is recommended. Values above this baseline indicate a higher-risk environment that requires basic safety measures, such as toolbox meetings. For the detailed index (i.e., ${N-SCR}_{L2}$), which reflects specific activity–environment interactions, a critical threshold of 2.4 (derived from the upper quartile of the high-density case) is suggested. Days when ${N-SCR}_{L2}$ exceeds 2.4 should be considered high-risk periods requiring close supervision or potential work stoppages. This hierarchical threshold system helps safety managers use lower-level indices for general awareness and higher-level indices for focused interventions.

6. Conclusions

This study proposes a quantitative methodology for assessing ADW risk by developing hierarchical indices based on ARM metrics. Unlike subjective or frequency-based techniques, this automated method provides a project-specific evaluation by analyzing concurrent environmental and work-related conditions.

The conversion of ARM metrics into hierarchical ADW risk indices (i.e., ECRs and SCRs) offers significant theoretical and practical implications. Theoretically, this research bridges the gap between static accident data and dynamic risk assessment. While previous studies using ARM primarily focused on identifying qualitative accident patterns, this study advances the methodology by converting ARM metrics into hierarchical quantitative indices. This formalization provides a theoretical framework for measuring interactions between environmental conditions and specific work activities, moving beyond simple frequency-based risk analysis. Practically, the method serves as an automated decision-support tool. By integrating ADW risk indices with the project’s WBS, the system enables daily and weekly ‘look-ahead’ safety planning, allowing practitioners to identify high-risk periods and allocate resources based on data-driven forecasts rather than intuition.

The case applications validated the method’s practical relevance. The indices successfully captured both expected and counterintuitive patterns: sharp increases in risk during trade overlaps and extreme risks in foundation works under harsh winter conditions, while interior works exhibited consistently lower risks under adverse conditions. These results demonstrate the method’s ability to detect context-specific fluctuations in ADW risk. Regarding generalizability, while this study focused on the building sector to ensure statistical significance, the core methodology—deriving risk from interactions between environmental conditions and schedules—is domain-independent. Therefore, given sufficient accident records, this framework can be effectively adapted to other sectors, such as civil engineering or industrial plant projects.

However, limitations remain. First, the current model lacks spatial attributes, such as site layout, pedestrian routes, and proximity to moving equipment, limiting the assessment of spatially dynamic hazards. Second, designed primarily for look-ahead planning, the framework does not account for transient site hazards, such as dynamic material storages. To address these limitations, future studies should aim to integrate the ADW risk assessment framework with BIM and Geographic Information Systems. This integration would foster the development of a more comprehensive and adaptive safety management system that accurately reflects the temporal and spatial dynamics of walking-related safety hazards on construction sites.