Understanding Artificial Neural Networks as a Transformative Approach to Construction Risk Management: A Systematic Literature Review

Abstract

The construction industry is characterized by complexity and high risk, making effective risk management essential for project success. Traditional risk management methods, which often rely on expert judgment and historical data, are increasingly inadequate for addressing modern construction projects’ dynamic and multifaceted challenges. This study systematically reviews applications of artificial neural networks (ANNs) in construction risk management, covering studies published between 1990 and 2024. Following PRISMA 2020 guidelines, an initial TITLE-ABSTRACT-KEYWORD search in Scopus (1990–2024) yielded 4648 records. After applying subject area and publication-type filters, 2483 records remained. Following duplicate removal, title and abstract screening reduced the pool to 132. After a full-text eligibility assessment, 86 studies were retained. Two additional studies were identified through co-citation analysis, and after the exclusion of four retracted papers, 84 studies were included in the final synthesis. Relevant peer-reviewed studies were categorized to evaluate ANN models, their applications, and key findings. The results indicate that ANNs, including backpropagation and radial basis function networks, have been applied effectively in cost estimation, schedule prediction, safety assessment, and quality control tasks. They offer advantages compared with conventional approaches, such as improved pattern recognition, faster data processing, and more accurate risk evaluation. At the same time, critical challenges persist, including data quality, computational demands, and the interpretability of outputs. To address these issues, studies increasingly recommend integrating ANNs with hybrid approaches such as fuzzy logic, genetic algorithms, and Monte Carlo simulations, as well as leveraging real-time data through IoT and BIM frameworks. This review contributes to theory and practice by consolidating fragmented evidence, distinguishing theoretical and practical contributions, and offering practical recommendations for industry adoption. It also highlights future research directions, particularly the integration of hybrid models, explainable AI, and real-time data environments.

1. Introduction

The construction industry is characterized by its complexity and susceptibility to risks that can significantly influence project outcomes. Traditional risk management approaches, which often rely on expert judgment and historical data, have been critiqued for their limitations in addressing projects’ dynamic and multidimensional nature [1,2]. Artificial Neural Networks (ANNs), which can model nonlinear relationships and learn from data, have been increasingly explored as a potential tool to enhance risk management practices in this domain.

Risk management holds particular significance in emerging economies, where rapid urbanization, limited resources, and inadequate infrastructure magnify the consequences of project risks. In these contexts, risk practices influence not only the outcomes of individual projects but also broader economic stability, employment, and social development. Recent evidence shows that adopting structured approaches to risk management can strengthen sustainable project performance, particularly when combined with stakeholder engagement as a mediating factor [3]. At the global level, the Global Construction Risk Review 2025 reports that contractors, especially in emerging markets, face disproportionate vulnerabilities such as financial volatility, inflation, supply chain disruptions, and climate-related risks, which collectively heighten exposure to construction-related challenges [4]. Complementing this perspective, another study highlights that in the Middle East and North Africa region, reliance on informal practices and subjective judgment often undermines effective risk assessment, thereby limiting resilience [5]. Collectively, these insights confirm that without adaptive and formalized strategies, construction projects in emerging economies risk intensifying systemic vulnerabilities rather than reducing them.

Technological advancements have catalyzed an intensification of data volume within the construction industry, an industry traditionally characterized by extensive information requirements. The proprietary information embedded within these datasets proves instrumental in formulating efficacious construction methodologies conducive to enhanced project efficacy. Consequently, the construction industry necessitates the adoption of advanced intelligent technologies, with Artificial Neural Networks (ANNs) being particularly well-suited to address the challenges associated with the accelerated pace of data generation in construction management.

Research on neural network models has predominantly concentrated on leveraging their capabilities to address nonlinear challenges. Nevertheless, inherent uncertainties and subjective factors may compromise the precision of estimation outcomes. Furthermore, the amalgamation of neural networks with complementary techniques holds promise for creating more efficient models. The exploration of neural networks in construction project risk management remains a vibrant scientific domain, as evidenced by the limited number of studies conducted over the past decade. ANNs have been employed in various risk management tasks within construction, such as cost estimation, project delay prediction, safety hazard identification, and quality control [6].

This review, therefore, concentrates exclusively on ANNs. Compared with other advanced approaches such as support vector machines, decision trees, or ensemble methods, ANNs have been more extensively adopted in the construction risk management literature, particularly for problems involving uncertainty, nonlinear relationships, and multidimensional datasets. Their suitability for modeling complex and dynamic risk interactions has made them one of this domain’s most commonly used AI techniques. At the same time, systematic reviews focusing specifically on ANN applications remain relatively limited, as most previous studies examined artificial intelligence or machine learning methods more broadly. By concentrating on ANNs, this study offers a targeted synthesis that consolidates their applications, identifies their advantages and limitations, and outlines methodological and practical development opportunities.

Given the constraints on resources available to construction management and production teams, it is crucial to accurately identify risk factors that significantly impact construction duration and costs throughout the project’s implementation [7]. While widely recognized as pattern recognition systems in multidisciplinary literature, neural networks can be readily adapted to address diverse problem types. With minimal effort, decision-making scenarios can be reframed as pattern recognition challenges.

Despite the growing interest in ANNs, there remains a lack of comprehensive understanding regarding their application in construction risk management. Existing studies have primarily focused on specific aspects of ANN implementation, leaving a gap in synthesizing these findings to provide a holistic perspective for researchers and practitioners [8,9]. Furthermore, the construction industry faces challenges such as the interdependence of risk factors, the evolving nature of project environments, and the limitations of traditional analytical methods, which necessitate innovative risk assessment and mitigation approaches.

This study seeks to systematically examine Artificial Neural Network (ANN) applications in construction risk management through a structured methodological approach. The research involves identifying relevant keywords, performing a focused search within the Scopus database—selected for its coverage of peer-reviewed scholarly literature—and applying appropriate screening criteria to extract data from the selected studies. By employing this methodological framework, the study aims to thoroughly analyze how ANNs are utilized in construction risk management contexts, contributing to a more cohesive understanding of current practices and potential developments in this field.

The objectives of the study include mapping the current state of ANN implementation in construction risk management, categorizing applications based on risk types and management approaches, and evaluating the effectiveness and limitations of different ANN techniques [10]. Additionally, the study synthesizes findings to guide future studies for selecting appropriate ANN techniques tailored to specific risk management needs. Practical guidelines for implementing ANN-based solutions are also discussed to bridge the gap between theoretical advancements and practical applications. The study systematically examines ANN-based risk management strategies, methodologies, applications, and performance. The research is guided by the key questions below that aim to uncover the current state of ANN implementation, evaluate their effectiveness, and identify opportunities for optimization and future development. Fundamental research questions form the foundation of the systematic literature review by serving as a conceptual lens through which the entire review process is conducted. These questions direct every stage of the methodology—from initial database searches and strict screening procedures to comprehensive data extraction and thematic synthesis—ensuring that the review remains centered on evaluating the transformative role of Artificial Neural Networks in construction risk management. This integrative approach facilitates a systematic examination of the efficacy, encountered challenges, and future potential of ANN-based approaches, reinforcing the study’s coherence and scholarly contribution. This study seeks to answer these questions: (1) How are artificial neural network (ANN) models reshaping the risk management landscape in construction, particularly in addressing emerging challenges beyond traditional methodologies? (2) What role do artificial neural networks play in enhancing risk management practices, including risk identification, assessment, prediction, and decision-making within construction projects? (3) How can improvements in data selection and validation enhance the predictive capability of ANN models in construction projects? (4) What future innovations and research avenues can further elevate the role of ANN in transforming construction risk management practices?

2. Materials and Methods

This study employs a literature review methodology to comprehensively analyze the application of Artificial Neural Networks (ANNs) in construction risk management. The systematic literature review methodology was chosen for its ability to provide a structured and comprehensive analysis of the existing body of knowledge. The construction industry faces increasingly complex risk management challenges that require sophisticated analytical approaches. ANNs have emerged as a promising tool in this context, yet their applications, limitations, and potential benefits remain fragmented across various studies.

The data collection process was conducted exclusively through the Scopus database, selected for its comprehensive coverage of peer-reviewed scholarly works in engineering and construction management. Selected papers had to meet at least one of the criteria to be included in the research. Publications were included if they were peer-reviewed journal articles or conference proceedings specifically focused on ANN applications in construction risk management. Conversely, publications were excluded if they were unrelated to construction risk management, lacked clear methodological descriptions, or were non-peer-reviewed materials. This systematic approach to literature selection helped maintain the focus and academic rigor of the review while ensuring that the analyzed studies contributed meaningfully to addressing the research questions.

The keywords were selected to ensure precision and relevance in capturing the literature. The terms “artificial neural networks,” “construction,” and “risk management” were combined to focus specifically on ANN applications within construction risk contexts, thereby excluding unrelated AI methods and non-construction domains. The review period of 1990–2024 was chosen because ANN applications in construction risk management began to gain visibility after 1990, while the cut-off at 2024 reflects the timing of this study and ensures that the most recent developments were included. This timeframe balances historical perspective with contemporary advancements, providing a comprehensive yet focused view of ANN applications in the field.

The review followed a structured methodological process to ensure transparency and reproducibility. First, Scopus was selected as the sole database for its comprehensive coverage of peer-reviewed works in engineering and construction management. The keywords “artificial neural networks,” “construction,” and “risk management” were applied to focus the search on ANN-based studies in construction risk contexts. The review period was set to 1990–2024 to capture early applications and the most recent developments.

The study selection process strictly followed PRISMA 2020 guidelines to ensure transparency and reproducibility. An initial TITLE-ABSTRACT-KEYWORD search in Scopus (1990–2024) identified 4648 records. After applying subject area filters (Engineering, Computer Science, Decision Sciences), publication type restrictions (journal articles and conference papers), and limiting to English-language records, the dataset was reduced to 2633 studies. Duplicate removal eliminated 150 records, leaving 2483 for screening. Title and abstract screening based on predefined inclusion and exclusion criteria narrowed the pool to 132 studies. Full-text eligibility assessment excluded 50 studies that lacked relevance to ANN applications in construction risk management, leaving 82 studies. Two additional studies were identified through co-citation analysis, while four retracted records were excluded, resulting in 84 studies for final synthesis. This process is documented in the PRISMA 2020 flow diagram (Figure 1). VOSviewer 1.6.16 was selected for the co-citation analysis because it is one of the most widely used and recognized tools for bibliometric mapping. It offers strong capabilities for constructing and visualizing bibliometric networks, handles large datasets efficiently, and provides reliable clustering functions. Compared with alternative tools such as CiteSpace or Bibliometrix, VOSviewer combines methodological rigor with user-friendly visualization, which makes it particularly suitable for this review. Finally, data extraction and thematic synthesis were conducted, classifying ANN architectures, application areas, performance metrics, and validation methods, and identifying key themes, contributions, and research gaps. To complement the PRISMA diagram, a research flowchart of the study (Figure 2) was added to visually link the identified research gap with the methodological steps of this study.

This systematic review followed the PRISMA 2020 guidelines. No prior protocol was registered, as registration is not mandatory for all systematic reviews. However, we ensured transparency by documenting the inclusion/exclusion criteria, search strategy, and study selection process. The PRISMA 2020 checklist was followed throughout the process to ensure transparency and reproducibility. Although no protocol was registered in advance, all steps of the search and screening procedure are reported in detail to comply with PRISMA standards. To ensure transparency and reproducibility, the study selection process is presented in a PRISMA 2020-compliant flow diagram (Figure 1), which details the number of records identified, screened, excluded (with reasons), and included in the final synthesis.

In addition, flowchart of the study has been included to clearly communicate both the research gap addressed and the process followed in this study (Figure 2). The upper part of the diagram highlights the research gap: while various AI methods have been applied in construction risk management, systematic reviews focusing specifically on ANN-based approaches remain relatively limited. The lower part of the diagram illustrates the methodological process of this review, including the formulation of research questions, keyword selection, database search in Scopus, screening and eligibility assessment in line with PRISMA, data extraction, and thematic synthesis. This diagram complements the PRISMA flow by visually linking the motivation for the study with the structured steps undertaken.

3. Results

3.1. Analytics of the Literature Review

A total of 132 papers were initially considered, with articles focusing on the medicine and health industry excluded from the research. After applying the inclusion criteria, 86 articles were retained for co-citation analysis, conducted using VosViewer. Following the eligibility assessment of 86 studies, an additional two studies were identified through co-citation analysis and included in the final synthesis, bringing the total number of included studies to 88. However, four of these studies were later excluded due to retraction notices, resulting in a final set of 84 studies included in the systematic review. The asterisk * serves as a wildcard operator, automatically including all variations that begin with the root term ‘risk’ (e.g., risk, risks, risk management, risk observation, risk analysis). The search strategy employed the following Boolean combinations, and the whole Scopus search string is given below:

TITLE-ABS-KEY ((“neural network” OR “artificial neural network”) AND construction OR building OR project* AND (“risk*”))

The systematic review reveals compelling patterns in adopting and evolving Artificial Neural Networks (ANNs) in construction risk management. Figure 3 illustrates the annual distribution of studies and their corresponding citation counts from 1992 to 2024. The blue bars represent the number of studies published each year, while the red dashed line with markers indicates the citation counts for those studies. Citation counts are annotated above the markers for clarity. The data highlights significant peaks in both study counts and citations, such as the high citation count in 2011 (212 citations) and the increased number of studies in recent years (2021–2022). The figure also reflects the expected lag in citations for recent publications (2023–2024), as these studies are too new to have accumulated substantial citations.

3.2. Temporal Evolution of ANN Applications

Figure 4 provides a chronological overview of the evolution of ANN applications in construction risk management, segmented into five distinct periods: the Foundation and Early Integration Period (1990–2005), Advanced Methods (2006–2010), Deep Learning Era (2011–2015), and AI-Driven Systems (2016–Present). Each period is characterized by specific themes, methods, and applications, reflecting the progressive sophistication of ANN methodologies, which is shown in Figure 4.

The VOSviewer visualization in Figure 5 presents a keyword co-occurrence network illuminating the intellectual landscape of artificial neural network applications in risk management and construction domains. The network reveals several distinct clusters, each representing thematic concentrations in the literature. The red cluster, dominated by terms such as “BP neural network,” “risk evaluation,” and “analytic hierarchy process,” demonstrates the prevalence of backpropagation neural networks integrated with multi-criteria decision-making techniques for risk assessment purposes. The green cluster encompasses keywords including “risk management,” “construction project,” and “artificial neural networks,” indicating substantial research dedicated to neural network applications within construction risk management contexts. The blue cluster, featuring terms like “neural network,” “risk analysis,” and “project management,” represents research focused on risk analysis frameworks and their integration into project management practices. The network structure demonstrates the central role of the “BP neural network” as evidenced by its numerous connections to other keywords, indicating its foundational importance in this domain. The strong linkages between “fuzzy neural network,” “AHP,” and “risk evaluation” highlight a significant trend toward methodological hybridization, where researchers combine multiple computational and decision-making techniques to develop more robust risk assessment models. Keywords such as “machine learning,” “deep neural network,” and “comprehensive evaluation “suggest emerging research directions, indicating an evolution toward more sophisticated neural network architectures. This visualization effectively captures the field’s interdisciplinary nature, demonstrating how computational intelligence techniques are adapted to domain-specific risk management challenges, particularly in construction contexts, with a collective research effort focused on enhancing the accuracy and practical applicability of neural network-based risk evaluation models.

Figure 6 systematically maps Artificial Neural Network (ANN) applications in construction risk management, categorizing the research into three interconnected dimensions: application domains, risk assessment techniques, and neural network architecture. Application Domains illustrate the diverse contexts in which ANNs are applied, ranging from construction projects and infrastructure to environmental and industrial domains. The breadth of application domains demonstrates the adaptability of ANN-based approaches in addressing various challenges within the construction industry. Risk Assessment Techniques showcase the methodological approaches integrated with ANNs, including fuzzy logic, Monte Carlo simulations, and hybrid methods. The variety of techniques reflects the evolution of risk management practices toward more data-driven and predictive methodologies. Neural Network Architectures highlight the range of ANN architectures employed, from foundational models like feedforward and backpropagation networks to advanced systems such as deep neural networks and fuzzy neural networks. This diversity underscores the increasing sophistication of ANN applications in managing complex risk scenarios.

The central node represents the overarching theme, linking the three main dimensions mentioned above to highlight their interdependence in advancing risk management practices. Figure 6 also emphasizes the shift from traditional methods to advanced AI-driven techniques, showcasing the growing reliance on automated and data-centric systems for effective construction risk management. By organizing the information into these dimensions, the visualization provides a comprehensive overview of the research landscape, offering valuable insights for researchers and practitioners.

The timeline of concepts in ANN applications for construction risk management across four distinct periods from 1990 to the present is represented in Figure 7. The analysis encompasses 84 studies, with their distribution revealing significant temporal patterns: 25 studies (1990–2010) established the foundational concepts, followed by 15 studies (2011–2015) and 13 studies (2016–2020) that bridged traditional approaches with emerging technologies. The recent surge of 31 studies (2021–present) demonstrates the field’s contemporary vitality. The interconnected nodes, represented by dashed lines, depict the progressive transformation from basic neural network applications to sophisticated AI-driven solutions.

3.3. Application of ANNs in Construction Risk Management

Artificial Neural Networks (ANNs) have been applied in the construction industry for risk identification, assessment, and prediction. This section synthesizes findings from various studies, highlighting the evolution of ANN applications in construction risk management from 2006 to 2023.

3.3.1. Evolution of ANN Architectures in Risk Management

The development of ANN applications in construction risk management (

Table 1. Overview of representative ANN applications in construction risk management, drawn from selected studies published between 1990 and 2024.

Source	ANN Type	Application Area	Key Finding
[10]	Backpropagation Neural Network	High-cutting slope construction	Effectively identified and predicted safety risks, providing valuable insights for improving safety protocols and reducing the likelihood of accidents.
[12]		Construction projects	Accurately identified and predicted project risks. Provided a versatile tool for risk management across different types of construction projects.
[13]		Construction projects	Enhanced the reliability of risk predictions by learning from historical project data. Provided a robust alternative to traditional risk assessment methods.
[14]		Construction operations	Identified critical risk factors, including labor productivity, material availability, and site conditions. Simulated various risk scenarios to provide a detailed understanding of potential risks and their impacts.
[15]		Specialized highway maintenance projects	Identified and predicted critical safety risks by analyzing historical safety data and maintenance records. Provided effective risk mitigation strategies, enhancing safety management.
[16]		Real estate project risk evaluation	Confirmed the model’s effectiveness in reducing human subjectivity while providing a more reliable and scientific approach to risk assessment in real estate development enterprises.
[17]		Power network planning projects	Evaluated various risk factors such as load forecasting errors, equipment failures, and operational inefficiencies. Isolated critical risk factors that significantly impact project outcomes.
[18]		Wind power projects	Accurately assessed and predicted investment risks, including cost overruns and project delays.
[19]		Real estate project risk assessment	Successfully demonstrated the model’s superior applicability and accuracy while effectively eliminating subjective biases in the assessment process, providing a more reliable framework for investment decision-making.
[20]		Knowledge management risk early warning systems	Revealed the network’s exceptional self-adaptive and self-learning capabilities enable automatic performance adjustments based on environmental changes, making it particularly effective for dynamic risk environments.
[21]		Prefabricated projects	Accurately identified key risk factors contributing to schedule delays. Enabled better project scheduling and risk management.
[22]		Waste incineration power generation projects	Pinpointed critical risk factors such as technical failures, environmental impacts, and regulatory compliance issues. Reduced uncertainty in risk assessments by accurately predicting risks.
[23]		Construction project risk prediction	Achieved an output of 0.043 (vs. expert forecast of 0.045), demonstrating a balance of computational efficiency with practical accuracy.
[24]		Construction projects	Enhanced safety management by accurately predicting and mitigating safety risks, contributing to safer construction environments.
[25]		Thermal power projects	Identified critical factors influencing project success, providing highly accurate predictions and validating its application in managing thermal power project risks.
[26]		Chemical projects	Effectively identified significant risk factors, enhancing risk management strategies in chemical engineering projects.
[27]		Highway tunnel portals	High accuracy in risk identification and prediction demonstrated its practical application in infrastructure projects, contributing to better risk management.
[28]	Deep learning neural networks	Offshore construction projects	Demonstrated high accuracy in predicting key risk factors, showcasing the effectiveness of deep learning techniques in addressing complex operational risks.
[29]		Road construction projects	Effectively identified key risk factors and provided a comprehensive framework for risk management, improving the accuracy of risk predictions.
[30]	Fuzzy neural network	Energy Performance Contracting project risk evaluation	Proved effectiveness not only in assessing EPC project risks but also demonstrated potential applicability across various EPC projects, particularly in mature Western markets.
[31]		Green building projects	Incorporated data from multiple sources, such as environmental impact assessments and regulatory frameworks, to provide comprehensive risk assessments. Improved the accuracy and reliability of risk predictions.
[32]		Prefabricated building construction	Provided a detailed assessment of the advantages and challenges associated with prefabrication, identifying critical factors such as supply chain management, construction quality, and cost control.
[33]		International construction projects	Focused on compliance and contract issues, providing enhanced accuracy in identifying regulatory risks and improving legal risk management.
[34]	GA based BP neural network	Project portfolios	Demonstrated high accuracy in risk predictions, providing valuable insights for managing project portfolios more effectively.
[35]	GCPSO-based ANN classifier	project risk assessment	Optimized ANN parameters with a PSO variant (GCPSO), outperforming traditional BP networks in accuracy and reducing training iterations.
[36]	Multilayer perceptron neural network	Seismic vulnerability assessment	Achieved impressive accuracy rates of 65.60% with the MLP Classifier and 67.82% with the Keras model, representing a significant improvement over traditional RVS methods.
[37]		High-rise construction projects	Effectively identified structural and logistical risks by analyzing data from similar projects. Allowed for more accurate risk identification and better risk management strategies in high-rise building construction.
[38]	Neural network with fuzzy interface	Software industry projects	Improved risk identification and prediction accuracy, helping project managers address potential issues and improve project outcomes.
[39]	Radial basis function neural networks	Butterfly arch girder bridges	Effectively identified key structural risks, contributing to better structural integrity management and improving overall project safety.
[40]		Real estate development risk evaluation	Developed an innovative evaluation system that effectively captured risk factors while providing a comprehensive index system that significantly improved risk assessment capabilities in real estate development projects.
[41]	LTSM neural network	Highway construction projects	Accurately identified potential scheduling risks, facilitating better project scheduling and risk management.

) shows a clear progression from simple architectures to more sophisticated models. Early implementations on multiple-dimension neural networks focused on essential risk identification in highway projects [8]. This evolved into more complex applications such as backpropagation (BP) neural networks [9] and hybrid models combining ANNs with fuzzy logic and support vector machines. Other studies further advanced these approaches by applying BP neural networks to high-cutting slope construction to identify and predict safety risks, offering valuable insights for improving safety protocols [10]. In addition, ANN models in highway construction were employed to identify and prioritize key risk factors such as traffic flow disruptions, environmental impacts, and construction delays, thereby improving the overall risk management framework for highway projects [11].

3.3.2. Risk Identification Methods

Neural networks have been the most frequently employed for risk identification in construction projects. Applications range from energy-related contexts such as waste-to-energy [22], wind power [18], and hydropower projects [23], to building and infrastructure works [13,42]. These models consistently highlighted productivity, material availability, and site conditions as critical risks, confirming the suitability of ANN for identifying key drivers of cost and schedule overruns [33,35].

Hybrid models have been increasingly developed to address the limitations of simple BP networks. Rough Set–BP frameworks [43] provided better classification of risk categories, while AHP–ANN models [44] balanced expert-based weighting with neural computation. Further examples include GA-optimized ANN systems [45], which improved reliability and convergence, and models integrating material [46] and human risk factors [47]. Enhanced ANN models have also been applied in lean construction supply chain risk management, providing improved frameworks for addressing operational uncertainties and supporting decision-making under complex conditions [48]. These studies demonstrate that hybridization strengthens the interpretability and robustness of ANN-based risk identification.

Environmental and sustainability-related risks were also integrated into ANN frameworks. Neural networks have been used to assess risks related to environmental factors and green construction practices [49], highlighting how ANN models adapt to new industry challenges. These applications broaden the scope of ANN-based risk identification, extending it from traditional cost and time factors toward sustainability and environmental performance.

Worker safety and ergonomics risks have received special attention in ANN research. Posture recognition models [50] enabled the early identification of ergonomic hazards, while safety risk early warning systems [10] applied ANN techniques to construction operations. These studies illustrate how ANNs extend beyond static risk registers to capture real-time safety threats dynamically on construction sites.

ANNs have also been applied in more specialized industrial settings, including power network planning [17], prefabricated building projects [32], and waste incineration [22]. Such applications show that ANNs can capture interdependent risks across diverse engineering domains, not only in conventional construction but also in industrial and energy systems where uncertainty is high [51].

Taken together, these studies reveal a clear trajectory. While early works relied primarily on BP neural networks [13,18,22,23,33,35,42], later contributions introduced hybrid approaches [43,44,45,46,47] to improve accuracy and reduce subjectivity. More recent research explores advanced deep learning and IoT-enabled ANN applications [10,50,51] for real-time and adaptive risk recognition. This progression demonstrates the increasing sophistication of ANN-based models in proactively identifying risks across diverse project contexts.

3.3.3. Risk Assessment Techniques

Artificial neural networks have been extensively employed to assess diverse categories of construction risk, consistently showing stronger predictive performance than conventional statistical methods. Early applications demonstrated that BP networks could capture nonlinear interactions among safety factors, for instance, in high-rise fire risk evaluation [52,53], crane operations [35,38], and accident risk analysis in tunnel and subway projects [33,40,41]. These studies collectively confirmed that ANNs provide more accurate and flexible assessment frameworks for equipment and safety-related hazards, compared with linear regression or rule-based approaches [14,15,21]. Beyond these domain-specific applications, models that integrate rough sets with RBF neural networks have also been developed to simplify network structures and enhance learning efficiency, providing project managers with more systematic tools for risk evaluation [54]. Similarly, ANN-based approaches have been extended to the safety assessment of green building construction, where optimized BP models improved reliability in identifying construction-related hazards [55].

Recent applications also demonstrate ANN’s capacity in highly specialized domains. For instance, backpropagation neural networks were successfully applied to bridge maintenance projects, outperforming regression techniques in risk scoring and classification [56]. Similarly, ANN-based models have been developed to predict tower crane safety risks in super high-rise construction, providing reliable early warnings for site management [57].

Environmental and sustainability-oriented risk assessments also benefited from ANN-based models. Neural networks were applied to evaluate certification risks in green building materials [36], energy efficiency in thermal power plants [37,39], hydropower station safety [17], and waste management systems [21,27]. Such studies emphasize the adaptability of ANNs to incorporate environmental variables, regulatory uncertainties, and operational complexity, often achieving higher accuracy than conventional multi-criteria methods [30,58]. Similarly, prefabrication and modular construction projects [21,32] have adopted ANN models to assess risks tied to off-site production and supply chain uncertainties, further broadening the application of neural networks to modern project delivery methods.

Hybrid approaches have been increasingly utilized to overcome the limitations of basic BP and RBF models. Rough Set–RBF frameworks [13] improved the classification of complex risk categories, while fuzzy-ANN systems [37,39] and ANN–Monte Carlo combinations [30,58] offered more robust assessments under uncertainty, reducing subjectivity in decision-making. Genetic algorithm-optimized neural networks [45] and other metaheuristic integrations [59] further enhanced accuracy and convergence, illustrating the benefits of combining machine learning with optimization techniques in risk assessment.

These methodological advances also meet the growing need for comprehensive assessment tools in high-risk contexts. Studies applying convolutional and recurrent neural networks [56] to accident prediction, or a multi-layer perceptron to assess geotechnical and structural stability [51], highlight the progressive move toward deep learning architectures. Such models can process large and unstructured datasets, enabling real-time and adaptive risk assessment for increasingly complex construction environments [60,61].

The literature demonstrates that ANN-based risk assessment methods consistently outperform traditional models by capturing nonlinear and interdependent risk factors across multiple domains. Applications span from safety risks in high-rise and underground works [33,35,38,40,41,52,53,54,55,56] to environmental and sustainability risks [17,22,27,30,36,37,39,58], and to innovative practices such as prefabrication [46,60,61]. The transition from simple BP networks [13,14,15,21] to hybrid [30,37,39,43,44,45,58,59] and deep learning approaches [51,56,60,61] reflect the ongoing refinement of ANN-based risk assessment, underscoring its central role in evaluating complex risks in construction projects.

3.3.4. Prediction Methodologies

Artificial neural networks have been extensively applied for risk prediction in construction, progressing from simple BP networks to advanced hybrid and deep learning architectures. Early implementations focused on forecasting cost overruns, delays, and safety hazards using BP and RBF networks [13,15,19,21,24,27,28,29,33,34,38,44]. These studies demonstrated that ANN-based prediction significantly outperforms linear regression by capturing complex nonlinear relationships among diverse project risks. Applications ranged from structural safety [19,24], fire and accident prediction [38,44], and high-rise buildings [15,21] to infrastructure and tunneling projects [13,33], showing the versatility of ANN methodologies across multiple construction domains.

The latest trajectory involves deep learning architectures, which have shown remarkable capacity to process unstructured and dynamic datasets. Convolutional neural networks (CNNs) have been applied for geospatial and visual risk analysis, while recurrent neural networks (RNNs) and long short-term memory (LSTM) models [60,61,62] supported real-time monitoring of project risks, including safety inspections, financial risks, and construction schedule forecasting. These models demonstrate superior adaptability compared with shallow architectures, as they can learn temporal dependencies and capture complex sequences of risk evolution across project phases.

To enhance predictive performance, optimization-based ANN models have become increasingly common. Genetic algorithm–optimized BP networks [30,31,32,58,59,63], particle swarm optimization (PSO) enhancements [30,32,64], and artificial fish swarm algorithms (AFSA) [58,65] improved convergence speed, prevented local minima, and yielded more reliable predictions in areas such as heating system risks [58], energy infrastructure [63], and complex scheduling problems [31,32]. These optimization-driven hybrids consistently provided higher accuracy than conventional ANN training, reinforcing the synergy between evolutionary computation and neural prediction in construction risk management.

Beyond optimization, hybridization with fuzzy logic and other decision-support frameworks has extended ANN’s predictive utility under uncertainty. Fuzzy ANN combinations [41,46,48,66] have been widely used for risks involving qualitative or incomplete data, such as prefabricated construction [46], project financing, and safety risk evaluation under uncertainty [41]. Monte Carlo–ANN hybrids [30,58] further enhanced predictive accuracy by integrating probabilistic simulations with neural computation, enabling better handling of uncertainty in cost and schedule predictions.

Collectively, prediction methodologies reveal a methodological evolution. Initial BP and RBF networks [13,15,19,21,24,27,28,29,33,34,38,44] laid the foundation by proving ANN’s predictive superiority over conventional approaches. Optimization-driven hybrids [30,31,32,58,59,63,64,65] significantly improved performance, while fuzzy and Monte Carlo integrations [30,41,46,48,58,66] enhanced robustness under uncertainty. Finally, deep learning frameworks [57,60,61,62,67] expanded ANN-based prediction into real-time and data-intensive domains. This progression underscores the central role of ANNs as predictive engines in construction risk management, evolving from simple forecasting tools into sophisticated, adaptive systems that inform proactive decision-making across cost, time, safety, and environmental risks. A consolidated overview of these studies, including risk identification, assessment, prediction, and key findings is presented in

Table 2. Overview of ANN studies addressing risk identification, assessment, and prediction in construction projects, based on selected studies from 1990 to 2024.

Source	Risk Identification	Risk Assessment	Risk Prediction	Key Finding
[10]	◉	◉	◉	Demonstrated high accuracy in assessing safety risks, predicting project timelines, and forecasting resource allocation using BP neural network models.
[13]	◉	◉	◉	Contributed to the field by developing predictive models aimed at assessing worker safety, predicting equipment failures, and monitoring site conditions.
[15]	◉	◉		Refined BP neural network methods for highway maintenance project safety risk assessment, providing reliable risk evaluation capabilities.
[18]	◉	◉	◉	Enhanced risk assessment by implementing BP neural networks in wind power projects, accurately assessing and predicting investment risks such as cost overruns and project delays.
[19]	◉	◉	◉	Utilized convolutional neural networks with geospatial information to predict non-residential building fire risks, achieving superior prediction performance.
[21]	◉	◉	◉	Implemented an SD-BP neural network model for predicting delays in prefabricated projects, successfully identifying and quantifying risk disruption effects.
[22]	◉	◉		Examined the use of backpropagation (BP) neural networks for risk assessment in waste incineration power generation projects, demonstrating improved accuracy and computational efficiency in identifying potential risks.
[24]			◉	Provided insights into real-time applications of BP neural network-based safety risk management systems for accident prevention protocols and risk mitigation strategies.
[27]		◉	◉	Demonstrated the integration of multiple data sources to achieve informed predictions regarding environmental impacts, potential cost overruns, and schedule delays.
[32]		◉	◉	Highlighted the ability of fuzzy neural network models to predict supply chain disruptions and quality control issues in prefabricated building construction.
[35]		◉		Combined Rough Sets with ANN to process fuzzy data in construction projects, reducing input nodes by 37% and shortening training time without sacrificing accuracy, though the ANN’s complexity remained a barrier.
[38]			◉	Explored neural networks for project risk and talent management prediction, effectively forecasting project success/failure rates based on personnel factors.
[41]	◉	◉	◉	Created an LSTM neural network model for risk-based construction inspection prioritization, effectively predicting requirement-based risks in highway projects.
[42]	◉	◉		Applied an ANN approach to assessing risks in infrastructure construction projects in Egypt, successfully identifying 157 risk factors across technical, project management, financial, environmental, and organizational dimensions.
[44]	◉	◉	◉	Proposed a hybrid AHP-ANN model for predicting major risks in Taiwanese construction projects, successfully identifying and quantifying various risk factors.
[45]	◉	◉		Applied factor analysis to extract common risk factors in Construction-Agent projects and fed them into a BP neural network, improving model transparency while relying heavily on subjective expert surveys.
[46]		◉	◉	Implemented a GA-BP neural network model for project portfolio risk prediction, demonstrating enhanced prediction accuracy through genetic algorithm optimization.
[48]	◉	◉		Advanced risk identification by applying neural networks to manage lean construction supply chain risks, outperforming traditional methods such as K-means and heuristic algorithms.
[50]	◉	◉		Explored the adaptation of recognition models for posture recognition using wearable sensors, validating incremental deep neural networks for ergonomics risk assessment for construction workers.
[52]	◉	◉		Developed a BP neural network model for high-rise building fire risk assessment, effectively guiding decision-makers in fire risk management.
[53]	◉	◉		Developed a multilevel BP neural network model for evaluating fire risks in high-rise buildings, demonstrating strong correlations between predictions and actual risk scenarios.
[54]	◉	◉		Proposed an RS-RBF model that integrated rough sets with neural networks, overcoming traditional limitations and enhancing project risk assessment capabilities.
[55]	◉	◉		Investigated green building materials certification risks using an improved LMBP model, achieving lower MSE, better gradient solutions, and improved training efficiency.
[56]		◉		Compared ANN and regression models for bridge risk assessment; the hybrid ANN-regression model achieved superior performance (RMSE: 1.8 vs. 2.3), emphasizing ANN’s capacity to capture nonlinear relationships.
[57]	◉		◉	Developed a back-propagation neural network for predicting safety risks of tower cranes in high-rise construction, achieving high accuracy in risk prediction scenarios.
[59]		◉		Evaluated highway investment risks using BP neural networks, achieving an MSE of 0.12; however, the small sample size (8 projects) limited the model’s generalizability.
[60]		◉	◉	Created a neural network-based evaluation model for project financing risks, effectively predicting financial risks in large-scale projects.
[63]		◉	◉	Designed an ANN-based early-warning system for IT outsourcing projects, achieving 92% accuracy in risk-level classification, though the system’s threshold determination struggled with project complexity.
[64]	◉	◉	◉	Combined rough sets with artificial neural networks for construction project risk prediction, demonstrating improved accuracy through hybrid modeling.
[65]	◉	◉		Optimized a BP neural network with the Artificial Fish Swarm Algorithm for smart heating projects, improving risk classification accuracy to 94.5% and addressing market, technical, and environmental factors.
[66]	◉	◉	◉	Developed a BP neural network model for managing scientific research project resource conflicts and predicting risks, showing high accuracy in resource allocation scenarios.
[67]	◉	◉		Trained an MLP model on 135 Algerian construction projects, achieving a correlation coefficient (R) of 0.88; the model prioritized strategic risks but faced regional generalizability constraints.
[68]	◉	◉		Developed a compensative fuzzy neural network for risk assessment in land consolidation projects, showing improved prediction capabilities through fuzzy logic integration.

.

3.4. Optimization of Risk Management Strategies

Artificial neural networks have increasingly been incorporated into optimization frameworks in construction risk management, moving beyond prediction to provide actionable strategies for cost, schedule, safety, and sustainability. ANN applications range from financial risk assessment with convolutional neural networks [60] to hybrid PSO–BP approaches in engineering projects [69], both of which enhanced prediction accuracy and stability while mitigating local minima challenges. Such hybrid and advanced models underscore ANN’s capacity to support optimized decision-making under complex and uncertain conditions.

Deep learning and recurrent models have further broadened optimization capacities. A DRNN–MSCA framework [48] delivered superior accuracy in construction risk identification compared to traditional ANN or GA-based approaches, while Elman networks proved effective in capturing temporal dependencies in power plant projects [64]. ANN systems integrating IoT data streams supported real-time risk management in port security [24], while BP networks applied to scientific research projects [70] improved predictions related to resource conflicts and time-sensitive risks. These studies collectively highlight how advanced ANN architectures adapt to dynamic environments and computationally intensive scenarios.

Hybrid ANN approaches have also proven valuable in addressing uncertainty. Fuzzy logic–ANN systems [8] facilitated cross-cultural risk identification in international projects, while Monte Carlo–ANN combinations [14] enhanced stochastic risk simulation for cost and schedule optimization. Expert system–ANN hybrids [71] improved consistency in project financing risk evaluation, and optimized BP models reduced error rates in software project risk assessment [72]. Sustainability has also been addressed through ANN-based credit risk frameworks [73], which integrated environmental variables into project risk assessment, reflecting the alignment of ANN-based optimization with eco-friendly priorities.

Resource allocation has emerged as another critical frontier. ANN-based models for multi-objective optimization have been proposed to guide resource distribution [74], and IoT-based ANN frameworks [75] improved real-time tracking, reducing waste and enabling adaptive allocation strategies. Similar approaches have been applied to optimize tunnel construction responses [27], automate safety risk management [43], and develop predictive maintenance systems [76]. A BP neural network–based evaluation model has also been applied to investment risk in electric power projects, showing reliable forecasting performance and reducing subjectivity in expert judgment [77]. Hybrid GA–ANN models [62] further extended optimization into BIM-enabled environments, balancing cost, time, and sustainability criteria.

These studies confirm that while ANNs are not inherently optimization algorithms, they function as powerful surrogate models within hybrid frameworks that incorporate GA, PSO, fuzzy logic, and Monte Carlo methods. By embedding ANNs in these optimization pipelines, researchers have improved predictive accuracy, real-time adaptability, and multi-criteria decision-making [8,14,24,27,43,48,60,62,64,69,70,71,72,73,74,75,76,77,78]. This methodological progression illustrates how ANN-based optimization has evolved from theoretical modeling into integrated, practical strategies that enhance safety, efficiency, and sustainability in construction risk management.

Table 3. Overview of ANN-based optimization approaches in construction risk management, summarizing implementation methodologies and key findings from selected studies (1990–2024).

Author (Year)	Optimization Approach and Methodologies	Application Domain	Key Findings and Impact	Implementation Method
[14]	Backpropagation Neural Network with Monte Carlo simulation	Risk scenario simulation	Provided robust uncertainty analysis through Monte Carlo integration	Integration of Monte Carlo simulation with backpropagation neural networks
[24]	BP Neural Network integrating real-time IoT data	Smart port construction risk management	Enabled dynamic risk monitoring for complex security scenarios	BP neural network-based system integrated with IoT data for real-time optimization
[27]	ANN-based optimization for risk response	Tunnel construction risk responses	Optimized risk response strategies for tunnel construction	ANN-based approach for automated safety response
[62]	Deep Recurrent NN with Modified Sine Cosine Algorithm (DRNN-MSCA)	Construction project risk management	Achieved superior accuracy by optimizing sensitivity data computation; adequate for quantifying construction risks	DRNN-MSCA, combining deep recurrent architectures with modified sine cosine optimization
[64]	Elman Neural Network	Power plant construction risk management	Captured temporal dependencies leading to improved risk prediction	An Elman network tailored for time-sensitive risk evaluation
[69]	PSO integrated with BP Neural Networks	Engineering project risk evaluation	Improved convergence speed and reduced local minima issues, leading to more reliable risk assessments	A hybrid approach combining PSO’s global optimization with BP neural network learning
[70]	Sophisticated BP Neural Network for resource conflict management	Scientific research project risk management	Improved prediction of time-related risks and resource conflicts in complex research environments	BP neural network assessing both tangible and intangible resource conflicts
[71]	Merged expert systems (ES) with ANN	Project financing risk evaluation	Improved decision-making consistency; required iterative expert feedback to refine training samples	Hybrid integration with iterative expert feedback
[72]	Optimized BP network weights and thresholds	Software project risk modeling	Reduced prediction errors by 18% compared to unoptimized networks	Weight and threshold optimization algorithms
[73]	Utilized BP networks with the incorporation of environmental factors	Green credit risk model in water conservancy projects	Achieved 91.7% test accuracy; highlighted challenges in standardizing eco-friendly metrics	Model development with environmental factor integration
[74]	Dynamic resource allocation systems	Large-scale project resource management	Enabled real-time resource tracking and allocation	Dynamic ANN-based systems integrated with IoT data
[75]	Dynamic resource allocation with IoT integration	Resource tracking and management in large-scale projects	Enabled real-time resource tracking and dynamic allocation	ANN-based dynamic allocation system integrated with IoT devices
[76]	Predictive maintenance strategies using ANNs	Overall safety and operational efficiency	Enhanced safety and operational efficiency through optimized predictive maintenance	ANN-based predictive maintenance models
[77]	Hybrid models for multi-criteria decision-making	Construction risk management	Enhanced decision-making by balancing competing objectives	Integration of hybrid models in risk evaluation
[78]	Fuzzy Neural Networks	Risk uncertainty modeling	Enhanced modeling of inherent uncertainties	Fuzzy neural network techniques

provides a concise overview of ANN-based optimization strategies, including applied methodologies, domains, and key outcomes. Table 3 also synthesizes significant contributions from recent studies, highlighting how neural networks have been integrated with techniques such as Monte Carlo simulation, fuzzy logic, and IoT-based real-time monitoring.

3.5. Applications of ANN-Based Risk Management Strategies

Implementing Artificial Neural Networks (ANNs) in construction risk management has demonstrated their transformative potential in real-world scenarios. By addressing complex risk factors such as cost, schedule, safety, and environmental concerns, ANNs have enabled construction projects to achieve higher efficiency, reduced risks, and improved decision-making. This section explores specific case studies and applications that validate the effectiveness of ANN-based optimization strategies, highlighting their impact across various domains.

3.5.1. Cost and Schedule Optimization

Artificial neural networks have been increasingly applied to optimize cost and schedule management in construction projects, consistently showing higher accuracy than traditional forecasting methods. Early approaches integrated ANN models with Monte Carlo simulations [14], achieving significant improvements in predictive accuracy, with reductions of 42% in schedule overruns and 22% in contingency reserves. These results highlight ANN’s capacity to support proactive budget control and resource allocation in large-scale projects. Complementary applications of BP networks further demonstrated their efficiency in predicting cost and time risks, reducing manual forecasting effort by 40% and validating outputs against expert assessments [23].

Extensions of ANN-based methods have targeted project-specific risks in specialized contexts. For instance, BP models trained in MATLAB were applied to evaluate licensing and technical maturity risks [79], successfully identifying high-risk factors with limited real-world validation. In infrastructure projects, ANN models identified 157 risk factors across five broad categories—technical, management, financial, environmental, and organizational- and demonstrated high accuracy in predicting cost variations [42]. Similarly, a schedule delay prediction system for prefabricated construction [21] leveraged an SD-BP neural network to accurately forecast delays, offering decision-makers an effective tool for managing time risks in modular building projects.

Advanced deep learning approaches have shown further improvements in prediction reliability. A DNN-based framework for road construction projects [29] significantly outperformed Bayesian and Random Forest models, achieving lower MAE values for schedule and cost forecasts. In the context of PPP projects, ANN models supported risk allocation decision-making [80], capturing nonlinear relationships among risk factors and producing satisfactory allocation outcomes. These examples illustrate ANN’s adaptability across traditional, modular, and partnership-based delivery systems.

Recent studies have also explored convolutional neural networks (CNNs) to predict financial risks within engineering project management [60]. By combining structured financial indicators with CNN-based modeling, these studies reported high efficiency and accuracy in cost-risk prediction, reinforcing the potential of deep learning to handle multidimensional and large-scale datasets.

Overall, the literature on cost and schedule optimization reveals a progression from BP and Monte Carlo integrations [14,23,79] to large-scale infrastructure and prefabrication-focused applications [21,42], and toward advanced deep learning frameworks [29,60,80]. Across these efforts, ANN-based models consistently outperformed alternative techniques, providing more accurate cost forecasts, reducing schedule delays, and offering practical decision-support tools for resource optimization in complex project environments.

3.5.2. Applications in Safety Management

Artificial neural networks have been increasingly applied to construction safety management, addressing fire hazards, equipment risks, transportation infrastructure safety, and real-time monitoring. Their application has consistently shown that ANN-based models outperform traditional safety assessment approaches regarding accuracy, efficiency, and adaptability.

In fire safety, BP neural networks have been successfully applied to assess risks in high-rise buildings [52,53]. These models provided reliable evaluations of fire safety ratings, reduced assessment time, and closely matched expert opinions, demonstrating ANN’s potential to enhance conventional fire safety protocols. Complementing BP-based approaches, convolutional neural networks (CNNs) combined with geospatial information further improved fire risk prediction for non-residential buildings [81], illustrating the advantages of integrating visual and contextual data into safety models.

Beyond fire hazards, ANNs have been applied to equipment and infrastructure safety. For example, a tower crane safety evaluation system based on ANN [57] achieved higher prediction accuracy and simulation reliability than traditional methods, offering practical decision support for safe crane operation. Similarly, ANN-based models have been applied to transportation infrastructure maintenance [15], where BP networks enabled accurate and reliable safety risk assessments across diverse highway maintenance projects.

ANNs have also advanced real-time and adaptive safety management. BP-based systems integrated with IoT technologies supported smart port safety monitoring [24], enhancing accuracy, generalization, and reliability in dynamic operational settings. In tunnel construction, ANN models optimized safety protocols through historical safety data analysis [27], reducing safety incidents by over 50% and significantly improving worker protection in high-risk underground environments. Deep learning–IoT hybrid systems [43] provided continuous construction site monitoring and early hazard identification, demonstrating the potential of ANN-based approaches for proactive risk mitigation in real time.

Taken together, the literature on ANN applications in safety management [15,24,27,43,52,53,57,81] demonstrates a methodological evolution: from BP-based fire risk assessments in high-rise buildings to CNN-enhanced fire prediction, from equipment and infrastructure safety to dynamic IoT-integrated systems. Across these domains, ANN-based approaches consistently improved prediction accuracy, reduced evaluation time, and enabled proactive hazard management, underscoring their transformative role in enhancing construction safety.

3.5.3. Environmental Risk Management

In sustainable construction practices, fuzzy neural networks have been effectively applied to manage environmental risks. Ref. [82] enhanced environmental risk management by integrating artificial neural networks (ANNs) into a DEMATEL-ANP framework, improving risk factor analysis through advanced pattern recognition. ANNs facilitated the identification of complex, nonlinear relationships among regulatory, supplier, market, and technical risks, refining the prioritization process. This hybrid approach strengthened decision-making by enhancing adaptability and accuracy, demonstrating the value of combining machine learning with multi-criteria decision-making techniques for more data-driven environmental risk assessments. This integration enables the project managers to identify and mitigate environmental risks early in the project lifecycle to avoid the likely adverse impacts on the construction process and the ecosystem. Furthermore, in green building projects, the application of ANNs is not only limited to scheduling but also to optimizing energy consumption and reducing carbon emissions. Additionally, ANN models have proved useful in managing flood risks in construction projects in flood-prone areas. These models provide accurate predictions of flood risks by analyzing weather data and historical flood patterns, and, therefore, project teams can implement adequate preventive measures and minimize potential damage.

3.5.4. Stakeholder Communication and Decision Support

Artificial neural networks have also contributed to optimizing stakeholder communication and organizational risk management by transforming how information is processed and shared across project participants. ANN-based decision support systems have been shown to significantly enhance communication efficiency, reduce delays in risk reporting, and improve stakeholder engagement [74]. These improvements illustrate the role of neural networks in bridging information gaps and supporting more transparent and timely decision-making processes in complex projects.

In addition to communication, ANN frameworks have been applied to real-time decision support systems. For instance, ANN-driven platforms provided stakeholders with dynamic and up-to-date project risk information [64], an essential feature in projects characterized by multiple stakeholders and complex risk interactions. Such systems not only improved situational awareness but also strengthened the adaptability of decision-making in rapidly changing project environments.

ANN-based approaches have also been employed to mitigate conflicts among stakeholders. Studies demonstrated that these systems could reduce conflicts by approximately one-quarter, enhancing collaboration and project harmony [75]. Similarly, ANN-driven tools have been designed to translate complex model outputs into stakeholder-friendly reports [45], making advanced risk analyses accessible and comprehensible. This transparency has proven crucial in improving trust, cooperation, and collective action in high-risk project scenarios.

These contributions [45,64,74,75] highlight how ANN-based systems support stakeholder engagement, conflict resolution, and transparent communication in construction projects. By improving the timeliness, reliability, and accessibility of risk information, ANNs enable more inclusive and collaborative risk management, strengthening the alignment between technical analysis and stakeholder decision-making.

3.6. ANN Approaches in Construction Risk Management

The evolution of construction risk management has shifted from expert-driven, judgment-based approaches to more sophisticated analytical methods, with artificial neural networks (ANNs) consistently emerging as a superior alternative. Unlike traditional techniques that rely on subjective experience and historical records, ANNs process large, multidimensional datasets, uncovering nonlinear patterns and interdependencies that conventional models frequently overlook. Across empirical studies, ANN-based approaches have substantially improved prediction accuracy, adaptability, and proactive risk mitigation, enhancing project outcomes in diverse construction domains [10,18,21,32,83].

Applications of ANN models in safety and technical risk management provide strong evidence of this superiority. For instance, BP networks achieved an accuracy rate of 85% in predicting safety hazards in construction projects by processing complex safety indicators such as weather conditions, worker behavior, equipment status, and site-specific hazards [10]. Similarly, ANN-based models demonstrated 20% higher accuracy in investment risk prediction, 35% improvement in technical risk assessment, and 25% better environmental risk evaluation in wind power projects [18]. Fuzzy ANN frameworks further advanced performance in prefabrication projects by improving component quality risk assessment by 30% and assembly risk prediction by 40%, while better handling uncertainties inherent to modular construction [32]. ANN-based methods have also proven highly adaptable in high-risk environments, with studies documenting their precision in supporting safety management strategies [83].

Technical and practical considerations are crucial for effective ANN implementation in construction. Data availability and accessibility remain central requirements. Large, structured datasets improve ANN performance significantly, while IoT-enabled data collection systems strengthen predictive reliability [29]. Studies in power network planning confirmed that when input data comprehensively reflect diverse project factors, BP neural networks provide more reliable evaluations [17]. At the same time, the quality of datasets is equally essential; preprocessing such as cleaning and normalization has been shown to increase ANN prediction accuracy by up to 30% compared with unprocessed data [18,32].

The computational requirements of ANN models represent another critical factor. High-performance computing infrastructure or cloud-based solutions are often necessary to manage large-scale applications [31]. While BP models can be trained on standard hardware [17], more complex architectures demand scalable computing resources. Nevertheless, these investments are offset by substantial benefits: ANN-based models have been shown to reduce project delays by up to 30%, generating significant cost savings in practice [21].

Despite these advantages, the “black box” nature of ANN models poses challenges for interpretability and stakeholder acceptance. Unlike transparent traditional methods, ANN outputs are often difficult to explain to non-technical users. To address this, researchers have proposed explainable AI frameworks designed to clarify how predictions are derived, improving transparency, trust, and adoption [32]. These approaches balance the technical complexity of ANN architectures with the practical requirement for stakeholder understanding and usability.

To summarize, the literature [10,17,18,21,29,31,32,83] indicates that ANN-based approaches have not only outperformed traditional methods in predictive accuracy and adaptability but have also spurred methodological advances addressing data, computational, and interpretability challenges. While technical considerations remain, the ongoing development of explainable, data-rich, and computationally efficient ANN frameworks positions these models as indispensable tools for proactive and adaptive risk management in construction projects.

3.7. The Evaluation of ANN Performance in Construction Risk Management

Evaluating artificial neural network (ANN) performance in construction risk management requires comprehensive metrics, robust validation techniques, and insights from real-world implementations. Across the literature, researchers have applied diverse evaluation criteria, ranging from error-based metrics such as Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE) to broader indicators of adaptability, sustainability, and long-term reliability [10,17,18,29,32,55]. Together, these studies highlight the need for systematic frameworks to capture both the short-term predictive accuracy and the long-term operational effectiveness of ANN-based systems.

Prediction accuracy remains a central benchmark. Studies have shown that ANN models achieve substantially higher accuracy than traditional methods, particularly in cost, schedule, and safety risk contexts. For example, BP networks demonstrated a 20% improvement in MAPE in wind power risk prediction [18], while safety risk models for construction achieved accuracy rates of 85% in hazard detection [10]. Similar improvements were documented in infrastructure risk assessments, where early ANN models prioritized accuracy and consistency indexes as key performance indicators [17].

Beyond accuracy, reliability and robustness are crucial measures of ANN effectiveness. Research on multidimensional ANN frameworks for highway projects reported a 40% improvement in stability compared to traditional methods [11]. More advanced studies broadened the performance assessment to include precision, recall, and related indicators [29], reflecting the shift toward more holistic evaluation strategies. Cross-validation has been widely applied to confirm ANN reliability, with prefabricated construction studies employing K-fold validation [21] and time-series validation [32] to demonstrate consistent performance across dynamic project conditions. Hybrid frameworks have also validated their robustness, with ANN–Monte Carlo models reporting low RMSE and MAPE values across hundreds of projects [56].

Real-world applications further confirm ANN performance advantages. ANN-based safety management systems in high-risk environments demonstrated a 45% improvement in hazard prediction [83], while cost–benefit analyses showed that ANN implementations reduced project delays by up to 30%, generating substantial savings [31]. In infrastructure and regional contexts, ANN validation using MSE, MAE, and correlation coefficients confirmed predictive accuracy and generalizability [67]. These outcomes demonstrate that ANN systems can provide measurable value beyond laboratory testing, especially when applied in operational construction environments.

Adaptability and sustainability metrics have also gained prominence. ANN models have proven capable of learning from new data and adjusting predictions in dynamic environments, particularly in road construction projects [29]. Studies emphasized the need for long-term adaptability by integrating rough set theory with RBF networks [55], enabling models to maintain accuracy despite changing project conditions. Resource efficiency has likewise emerged as a key sustainability measure, as large-scale ANN applications, though computationally intensive, ultimately provide cost-effective and scalable solutions [31].

Finally, long-term performance monitoring is essential to sustain ANN effectiveness. Continuous recalibration and updates have been shown to maintain and improve accuracy over time [15,83]. These findings suggest that ANN-based risk management systems require ongoing oversight but reward such efforts with performance improvements as high as 60% compared to baseline methods.

The evidence [10,11,15,17,18,21,29,31,32,55,56,67,83] demonstrates that ANN evaluation frameworks must address multiple dimensions: accuracy, robustness, adaptability, and sustainability. While metrics such as MAPE, RMSE, and MAE provide essential quantitative benchmarks, broader validation approaches and long-term monitoring ensure that ANN systems deliver consistent, scalable, and trustworthy risk management solutions across diverse construction environments.

4. Challenges Using ANNs in the Construction Industry

Applying ANNs in construction risk management presents several challenges that must be addressed for their broader adoption. A primary concern is the availability and quality of training data. ANN models require large, representative datasets to achieve accuracy and generalizability, yet construction projects often generate fragmented or inconsistent data. Variability in data quality can significantly affect model performance, as shown in safety risk assessments [10,18,21], infrastructure risk prediction [15], and multidimensional highway risk models [11]. Limited reliance on single case studies further restricts the external validity of ANN models, underscoring the need for more generalized frameworks that can be applied across different project types.

Another significant challenge relates to ensuring data reliability across specialized contexts. Deep learning approaches in offshore construction require robust operational datasets [28], while hybrid ANN systems for coastal projects depend on diverse environmental inputs [84]. Similarly, pipeline risk models need comprehensive material failure records [58], and international projects demand accurate regulatory datasets [33]. Recurrent neural networks used for scheduling optimization also require large-scale, time-dependent datasets, which are not always accessible in practice. The lack of standardized data collection protocols hinders model reproducibility and cross-project applicability.

Beyond data limitations, the interpretability of ANN models remains a critical barrier to practical use. The “black-box” nature of neural networks makes it difficult for stakeholders to understand how predictions are generated, limiting trust and acceptance compared with more transparent traditional methods. This challenge has been observed across various ANN applications, from GA-BP investment risk models [25,34] to fuzzy and RBF-based models in tunnel and bridge projects [27,39]. To address this, researchers have begun exploring explainable artificial intelligence (XAI) approaches, including SHAP and LIME techniques, which provide insights into feature contributions and attention-based mechanisms that improve model transparency. Incorporating such interpretability tools enhances stakeholder confidence and ensures ANN adoption in real-world project environments.

Another challenge relates to methodological threats to validity. Published studies may be subject to publication bias, with positive ANN results more likely to appear in the literature. There is also considerable heterogeneity in how “risk” is defined across studies, complicating synthesis. Many models rely on small datasets, raising concerns about overfitting and limiting generalizability. Furthermore, the external validity of ANN approaches across different project types and geographical contexts remains underexplored. Addressing these threats requires larger and more diverse datasets, standardized definitions of risk, and cross-project validation to ensure robustness.

Computational demands also pose a challenge. Training and deploying ANN models often require high-performance computing resources, particularly for deep or hybrid architectures [31]. While cloud-based and distributed systems can mitigate these issues, smaller organizations may find it difficult to invest in the required infrastructure. Model integration represents another barrier, as embedding ANN-based systems into existing project management workflows and risk management frameworks requires technical expertise, organizational readiness, and stakeholder alignment.

In summary, the main challenges of applying ANNs to construction risk management can be grouped into four categories:

• Data Availability and Quality: ANN models depend on large, high-quality, standardized datasets, which are often limited in construction contexts.
• Model Complexity and Interpretability: ANNs’ black-box nature limits transparency and stakeholder trust, requiring the adoption of explainable AI methods.
• Computational Demands: Effective training and deployment of ANN models necessitate considerable computing resources, which may not be feasible for all organizations.
• System Integration: Incorporating ANNs into existing project management and risk assessment frameworks remains complex and resource intensive.

5. Potential Future Research Directions and Improvements

Advancing the effectiveness of artificial neural networks (ANNs) in construction risk management requires a set of strategic research directions that respond both to current limitations and to emerging technological opportunities. One of the most urgent needs is the development of standardized and representative datasets [85]. Such datasets would allow for consistent training, benchmarking, and validation across different construction project contexts, ensuring generalizability and reproducibility. Simplifying ANN architectures remains an important avenue for improving accessibility to non-technical practitioners, while enhancing computational efficiency is essential for enabling real-time risk prediction and monitoring in large-scale and dynamic projects. Hybrid approaches that combine ANNs with conventional statistical or expert-based methods continue to show promise by leveraging the complementary strengths of each [83].

Domain-specific research also highlights critical needs. In offshore construction, efficient deep learning models are required to address high computational demands [28]. For coastal and environmental projects, hybrid ANN frameworks must integrate diverse environmental datasets to provide accurate predictions of ecological and sustainability-related risks [84]. Similarly, improved collection of material failure data can enhance ANN models in pipeline construction [58], while adaptive recurrent neural networks (RNNs) are needed to capture the dynamic nature of scheduling risks in large-scale projects. In international contexts, access to comprehensive regulatory datasets remains essential to model legal and compliance-related risks [33].

Expanding sample sizes and diversifying data sources represent another vital research direction. Increasing sample sizes for specialized applications such as tunnel portals [27] can improve reliability, while developing risk assessment models that account for macro and micro factors is crucial for complex structures such as butterfly arch continuous girder bridges. External variables should be integrated more systematically, as demonstrated by improvements in thermal power project risk assessments when BP networks accounted for additional external factors [86]. Applying GA-BP models across broader predictive contexts can enhance generalizability [34], while extending neural network applications beyond the software industry will provide more substantial evidence of their effectiveness in diverse sectors [87].

Emerging technologies also open new opportunities for ANN-based risk management research. Integrating Building Information Modeling (BIM) and Digital Twins can facilitate real-time data exchange between physical and virtual project environments, improving design coordination, clash detection, and proactive risk identification. Similarly, combining ANN frameworks with Internet of Things (IoT) sensors and cloud/edge computing infrastructures will enable continuous monitoring and adaptive decision-making in highly dynamic construction environments. Explainable AI (XAI) methods—such as SHAP and LIME—should also be incorporated to address the “black box” challenge, ensuring transparency and enhancing stakeholder trust in ANN predictions. Furthermore, embedding sustainability metrics directly into ANN-based risk management models is necessary to align with global construction priorities, including carbon reduction, energy efficiency, and resilience.

These research directions emphasize the dual need to overcome technical and methodological barriers while embracing Industry 4.0 opportunities. By addressing data quality, interpretability, scalability, and integration challenges, and by leveraging hybrid approaches, IoT, BIM, and Digital Twin technologies, future studies can unlock the full potential of ANNs. These advancements will enhance the robustness and adaptability of risk management frameworks and ensure that ANN-based systems contribute meaningfully to improving safety, cost efficiency, and sustainability outcomes in construction projects. In the near term, the most actionable research directions involve improving dataset standardization and developing explainable AI frameworks, as these directly address the most frequently observed limitations across the reviewed studies.

6. Conclusions

This review synthesized evidence from 84 Scopus-indexed studies, focusing on applying Artificial Neural Networks (ANNs) in construction risk management. The findings confirm that ANNs have significant potential in recognizing, analyzing, and mitigating risks across multiple aspects of construction projects. They consistently outperform traditional risk management tools by capturing nonlinear relationships, processing large datasets, and updating predictions as new data becomes available. Nevertheless, the review also highlights several limitations that must be addressed before ANN-based approaches can be widely adopted in practice.

A key conclusion is that integrating ANNs with other risk management tools and emerging technologies is critical to enhancing their performance. Hybrid frameworks that combine ANNs with genetic algorithms, fuzzy logic, or Monte Carlo simulations provide more robust and reliable risk predictions. Similarly, integration with Industry 4.0 technologies, including Building Information Modeling (BIM), Digital Twins, and Internet of Things (IoT) sensor networks, offers opportunities for real-time monitoring, proactive decision-making, and improved stakeholder engagement. These synergies represent a promising pathway for future risk management systems that are adaptive, data-driven, and sustainable.

At the same time, several challenges remain that limit the widespread adoption of ANN-based approaches in construction risk management. Model performance is highly dependent on data availability and quality, which are often fragmented or inconsistent across projects [10,18,21]. Some applications remain case-specific [15], raising concerns about generalizability to different project types. Computational demands and scalability issues also restrict their real-time application, particularly in resource-constrained organizations [31]. Most critically, the black box nature of ANNs constrains interpretability and stakeholder trust, making practical implementation more difficult without complementary explainable AI tools such as SHAP, LIME, or attention-based mechanisms. Integrating ANN models into existing project management workflows requires technical expertise, organizational readiness, and stakeholder engagement. These limitations highlight the importance of developing explainable, scalable, and data-rich frameworks to ensure ANN models can be successfully translated into practice.

This review contributes to the academic understanding of ANN applications in construction risk management by combining findings from a wide range of studies. It organizes previous research into thematic areas such as risk identification, assessment, prediction, optimization, and performance evaluation. By outlining methodological progress—from basic BP networks to hybrid and deep learning models—the review clarifies how ANN approaches have evolved and how they compare with traditional risk management techniques.

From a practical standpoint, the review highlights insights that can assist practitioners in applying ANN-based methods in construction projects. It discusses the importance of data quality and availability, identifies challenges related to computational resources and interpretability, and points to the need for explainable AI techniques to improve stakeholder trust. In addition, the review notes opportunities for integrating ANN approaches with tools such as BIM, Digital Twins, and IoT-based monitoring systems. These applications can support decision-making, enhance safety, and improve resource allocation in dynamic project environments.

In line with the stated objectives, this review mapped how ANNs have been applied in construction risk management for tasks such as risk identification, assessment, prediction, and optimization. It also clarified the main advantages and limitations of these models compared with traditional methods, and highlighted future research needs including hybridization, explainable AI, and links with Industry 4.0 tools. Beyond these aims, the review showed that ANN-based methods can support project delivery, cost control, and safety when backed by reliable data and expert input. At the same time, barriers such as the ‘black box’ nature of neural networks and the limited availability of high-quality datasets continue to restrict broader use. It is worth noting that in emerging economies, where urban growth and resource constraints increase the weight of construction risks, ANN applications could provide resilience and efficiency if combined with supportive policies and capacity building. Overall, the review offers theoretical insights for researchers and practical recommendations for both practitioners and policymakers, helping to connect methodological progress with real project environments.

This review addressed three main research questions. The first concern was how ANNs have been applied in construction risk management. The findings show that ANNs have been used in risk identification, assessment, prediction, optimization, and performance evaluation across diverse project contexts. The second question asked what advantages and limitations ANNs present compared with traditional methods. The results demonstrate superior accuracy and adaptability but highlight challenges related to data requirements, computational demands, and interpretability. The third question explored future directions for ANN research in construction. Here, the literature points to hybrid models, real-time applications, and integration with broader frameworks such as BIM, Digital Twins, IoT, and explainable AI as important areas for future development.

For practitioners, several practical recommendations can be drawn from this review. First, efforts should be made to collect and maintain standardized datasets to improve the reliability of ANN applications. Second, organizations should consider hybrid ANN models that integrate complementary techniques such as genetic algorithms, fuzzy logic, or Monte Carlo simulation to improve robustness. Third, integrating ANN-based systems with digital tools such as BIM, IoT-enabled monitoring, and Digital Twins can provide real-time insights and support adaptive decision-making. Finally, adopting explainable AI methods will ensure that ANN outputs are transparent and trusted by all stakeholders involved in construction risk management.

Despite these contributions, this review has certain limitations. The search was restricted to the Scopus database, which may have excluded relevant studies indexed elsewhere. In addition, no formal risk-of-bias or study quality appraisal framework was applied, which may limit the assessment of methodological rigor in the included studies. These constraints should be considered when interpreting the findings, and future reviews could address them by adopting multi-database searches, broader search strategies, and standardized quality appraisal protocols.

Overall, ANNs represent a valuable addition to the toolbox of construction risk management, providing enhanced accuracy, adaptability, and efficiency compared with conventional methods. Yet their full potential will only be realized when challenges of data quality, computational demand, and interpretability are adequately addressed, and when integration with Industry 4.0 technologies is more systematically pursued. By combining technical improvements with practical implementation frameworks, ANN-based systems can evolve from experimental applications into core components of resilient, sustainable, and intelligent risk management in the construction industry.