New Trends in the Use of Artificial Intelligence and Natural Language Processing to Occupational Risks Prevention

1. Introduction

Occupational Safety and Health (OSH) is a major concern for all countries worldwide (Zhang et al., 2019), and its management is an ongoing challenge to protect the health and safety of workers and ensure a safe and healthy working environment. The latest International Labor Organization (ILO) estimates that nearly three million workers die each year from work-related accidents and diseases, an increase of more than 5 per cent compared to 2015 (ILO, 2023), denoting an urgent need for more action to prevent work-related accidents and diseases. Although the statistics remain alarming, working conditions have improved tremendously over the years (Badri et al., 2018). Advances in science and technology, such as engineering controls, safer machinery and processes, collective and individual protective equipment, and the implementation of regulations and labor inspections, have significantly reduced the incidence of occupational accidents and diseases associated with industrialization (Kim et al., 2016). In addition, the development of a preventive culture in organizational settings have been crucial in optimizing health and safety management. Industrial development began worldwide in the eighteenth century (Sharma & Singh, 2020) and has been characterized by a series of events that have triggered different revolutions over the years, as shown in Figure 1. These revolutions have been driven by technological transformations that have led to changes in the way industry operate and important social changes (Vinitha et al., 2020; Xu et al., 2018).

The First Industrial Revolution began with the introduction of the power loom in 1784 and was characterized by shifts towards the intensification of work activities. In this period, water power and the steam engine played a decisive role, both in their contributions to industry and transport (Sharma & Singh, 2020). In the 19th century, the Second Industrial Revolution was born with the invention of electricity production, innovations in development, the use of new materials (alloys, synthetic plastics), mass production and assembly lines (Zhang & Yang, 2020). With the appearance of the first programmable logic controller (PLC) in 1969, the transition from the invention and manufacture of analogue to digital electronic devices, automation, and the incorporation of information technologies (ICT) into industrial processes, the Third Industrial Revolution was born (Fonseca, 2018), which encouraged the glocalization of production and the relocation of jobs (Roberts, 2015).

The Fourth Industrial Revolution, or Industry 4.0 (early 21st century), is marked by technological developments with a certain autonomy and self-behavior, mainly focused on industrial automation robotic production through the integration of digital technology, information and communication technologies within an intelligent environment (Leesakul et al., 2022; Milea & Cioca, 2024; Gomez Miranda and Gonçalves, 2024). Specifically, this digital transformation is driven by technologies such as Blockchain, the Internet of Things, Big Data, Cyber-physical systems, Cobotics, Artificial Intelligence (AI), Natural Language Processing (NLP), Cloud Computing, Augmented Reality (AR) and Virtual Reality (VR), which aim to optimize production processes, increasing productivity and efficiency(Milea & Cioca, 2024). Finally, from 2021 onwards, the futuristic Fifth Industrial Revolution emerged, based on humans-robot collaboration to increase creativity and innovation by allowing robots to perform monotonous activities (Miraz et al., 2022).

On the other hand, the evolution in the field of OSH has always followed the different revolutionary advances in the industry (Badri et al., 2018) which has made it possible to react and propose effective solutions to be able to control occupational risks that may manifest themselves in the face of technological advances, innovations, changes in working methods, organization, work teams, processes, products, and workplace itself. Indeed, the nature of a changing work environment brings with it a series of OSH challenges and opportunities, which is why its management is essential to ensure workers' health, business sustainability and social stability (Wang et al., 2020)

Over the years, in most industrialized and developed countries, reactivity has given way to proactivity (Badri et al., 2018) which promotes a fully preventive approach to OSH that allows action from the source to eliminate risks. This allows the appropriate decisions to be made sufficiently in advance to anticipate possible undesirable events that could harm workers. To this end, the incorporation of digitalization is now beginning to offer new opportunities to innovate, improve and address new and emerging risks in the field of occupational risk prevention, through the incorporation of neurocognitive computing technologies, AI and NLP. Recent studies highlight the role of AI in the risk of occupational disease, analyzing workplace hazards, and enhancing safety measures (Garvin & Kimbleton, 2021; Howard & Schulte, 2024; Mollaei et al., 2023; Pishgar et al., 2021; Westhoven, 2022; Yimyam & Ketcham, 2022).

Thus, the evolution of occupational risk prevention has been characterized by a continuous expansion of its focus and methods. Initially focused on occupational health, it has transformed into a comprehensive approach that includes safety and health management and the early use of emerging technologies for the benefit of workers. This shift reflects a deeper understanding of workplace hazards, including chemical risks and psychosocial factors. The development of a preventive culture within organizational settings has been crucial in optimizing safety and health management systems. Recent strategies include dynamic risk assessment which helps organizations become better able to adapt to rapidly changing business or technological dynamics, putting them in a better position to respond to changes in business processes and their associated OSH risks (van Gulijk, 2021).

Today, this evolution is experiencing breakthrough, especially with the integration of AI and NLP. Recent studies highlight the role of AI in assessing the risk of occupational disease risk, analyzing workplace hazards, and enhancing safety measures (Garvin & Kimbleton, 2021; Mollaei et al., 2023; Pishgar et al., 2021; Westhoven, 2022; Yimyam & Ketcham, 2022) Specifically, NLP is demonstrating great potential in processing and interpreting large data sets for risk analysis. These technologies are central to the development of more efficient, accurate, and predictive occupational risk prevention strategies.

The objective of this study is to conduct a systematic review of the application of Artificial Intelligence (AI) models, particularly Large Language Models (LLMs) and Natural Language Processing (NLP), within the domain of occupational risk prevention. This review aims to elucidate how advanced technologies are being employed across various industrial sectors, including aviation, construction, the chemical industry, and transportation, to improve workplace safety and risk management. Specifically, the study will identify key technological applications such as real-time risk mapping, automated safety incident classification, and predictive modeling of occupational hazards. Furthermore, it seeks to address current challenges related to data quality, model transparency, and multilingual support while providing insights for future research to overcome these limitations and advance the efficacy of AI-driven occupational risk prevention strategies.

Recent work illustrates how AI and, in particular, large language models are reshaping occupational safety along the whole data–decision pipeline. At the level of risk assessment guidance, Baek et al. develop a retrieval-augmented LLM that mines 64,740 construction accident reports to automatically generate task- and equipment-specific safety risk management guidance, achieving quality comparable to experienced practitioners and reducing supervisors’ workload (Baek et al., 2025). Bernardi et al. extend this idea by combining RAG with explainable LLMs and layer-wise relevance propagation to produce job safety reports from unstructured aviation incident narratives, explicitly highlighting root causes to support human-centred, auditable decision-making (Bernardi et al., 2025).

A second cluster of studies focuses on structuring occupational information and safety data. Kim et al. fine-tune DistilKoBERT on nearly 100,000 survey responses to classify workers into 142 occupational codes with >84% accuracy, enabling large-scale, text-based occupational epidemiology (Kim et al., 2024). In parallel, Li et al. propose LLM4Jobs, an unsupervised framework that uses LLM-based summarization and embeddings to map noisy job descriptions to standard taxonomies such as ISCO/ESCO, outperforming prior unsupervised occupation coding methods (Li et al., 2025). Song et al. apply KoBERT to industrial accident descriptions to automatically classify occurrence types, mitigating subjectivity and inconsistency in manual coding and improving the quality of national accident statistics (Song et al., 2024).

Several contributions target predictive modelling of accident severity from textual or proactive safety data. Khairuddin et al. design an optimized Bi-LSTM architecture that fuses TF-IDF and GloVe embeddings of OSHA injury narratives, achieving up to 0.98 accuracy for amputation prediction and revealing salient causal keywords (Khairuddin et al., 2024). Using proactive Fatality Risk Control Programme data from an Indian steel plant, Sarker et al. combine NLP features with an ensemble of six classifiers in a soft-voting scheme to predict potential accident severity, supporting earlier and less biased interventions (Sarker et al., 2025). At the critical end of the severity spectrum, Parikh et al. integrate transformer-based text encoders with XGBoost to automatically flag incident reports involving potential serious injuries and fatalities (PSIF), using weak labelling and F2-optimised tuning to prioritise recall of high-risk cases (Parikh et al., 2024).

Finally, new multi-modal and reasoning-oriented architectures broaden the scope of AI-enabled prevention. Chen et al. introduce Clip2Safety, a zero-shot vision–language framework that recognises scenes, detects PPE and verifies fine-grained attributes across six workplace scenarios, improving accuracy and inference speed over prior VLM baselines (Chen et al., 2025). In general aviation, Liu et al. show that HFACS-guided chain-of-thought prompting (HFACS-CoT and HFACS-CoT+) markedly improves GPT-4o’s ability to infer pilots’ unsafe acts and preconditions from witness narratives, in some cases matching or surpassing human experts and exemplifying how domain knowledge can structure LLM-based accident investigation (Liu et al., 2025). Together, these studies demonstrate a shift from purely reactive analysis toward proactive, interpretable and context-aware AI systems for occupational risk prevention.

The present review occupies a specific niche at the intersection of occupational safety and health (OSH), natural language processing (NLP), large language models (LLMs) and, more broadly, AI-enabled safety analytics. Previous reviews have synthesized AI applications for industrial safety or Industry 4.0/5.0 more generally, but they have either focused on traditional machine-learning models, structured process data, or sector-specific issues (e.g. aviation, construction or healthcare) without systematically addressing text-centric and language-based approaches to occupational risk prevention. In contrast, this review concentrates on models that operate directly on unstructured OSH information—accident narratives, inspection reports, medical records, work-condition surveys or safety guidelines—and on how these models can be coupled with risk assessment and decision-support frameworks across sectors such as construction, mining, chemical and process industries, and healthcare.

In addition, this review extends the temporal scope to include the most recent wave of Gen-AI and LLM-based approaches that were not covered in earlier syntheses. These include ensemble learning models that map accident narratives to potential accident severity and Fatality Risk Control Programme (FRCP) levels (Sarker et al., 2025), coal-mine accident risk analysis frameworks that couple LLM reasoning with Bayesian networks (Du & Chen, 2025), improved topic-model–Bayesian-network pipelines for chemical safety risk identification (Zhou et al., 2025), multi-source data-driven risk assessment systems for coal-mine environments (Lu et al., 2024), and process-safety indicators derived from Industrial Internet data and LLM-assisted retrospective analysis (Ni et al., 2024). Furthermore, we incorporate recent developments in industrial accident type classification using KoBERT (Song et al., 2024), occupation and job-code classification with DistilKoBERT and LLMs (Kim et al., 2024; Li et al., 2025), multi-source heterogeneous data integration for incident likelihood analysis (Kamil et al., 2024), AHP-based studies of human error and environmental factors in mine accidents (Bekal et al., 2024), optimized deep-learning models for injury-severity prediction (Khairuddin et al., 2024), automatic identification of potential serious injuries and fatalities (PSIF) (Parikh et al., 2024), LLM-based accident-investigation reasoning guided by HFACS (Liu et al., 2025), RAG-enhanced large-language-model frameworks for safety guidance and job safety reports (Baek et al., 2025; Bernardi et al., 2025), and vision–language models for PPE-compliance monitoring (Chen et al., 2024). By integrating these very recent contributions, the review provides an updated and explicitly text-centric state-of-the-art on AI, NLP and LLMs for occupational risk prevention.

Several recent surveys focus on broad AI/OSH trends, risk management or Industry 4.0, but only tangentially cover text-based AI, natural language processing (NLP) and large language models (LLMs). For instance, Pishgar et al. (2021) provide the REDECA framework and a bibliometric mapping of AI in OSH, but do not systematically analyse the specific contribution of NLP and LLMs for processing unstructured safety narratives. Wang et al. (2020) map research domains in occupational health and safety management but treat AI only as one of many emerging themes. More recently, Gomes-Miranda and Gonçalves (2024) examine Industry 4.0 technologies and OHS, with AI appearing mostly as a component of digital transformation rather than as a dedicated methodological focus.

In contrast, the present review specifically targets text-centric AI methods, including classical NLP, deep learning architectures (e.g., CNNs, Bi-LSTM, BERT-type transformers) and, more recently, generative LLMs and retrieval-augmented generation (RAG)—as applied to accident reports, near-miss narratives, safety inspection records, occupational disease risk assessment and related OSH documentation

2. Methodology

This review explores how artificial intelligence (AI) models, especially natural language processing (NLP) techniques and large language model (LLM)–based approaches, are being applied to occupational risk prevention across industrial sectors. The aim is to map the breadth and characteristics of current applications. Accordingly, the methodology follows a structured approach, with four main stages: (i) identification of records through database and citation searches, (ii) screening and eligibility assessment based on predefined criteria, (iii) selection of studies for inclusion, and (iv) structured data extraction and narrative synthesis.

The review includes AI methods applied to text-centric or text-enriched safety data, accident and incident reports, near-miss and hazard observations, investigation narratives, work and exposure records, safety-management documentation, and, in some cases, multi-source frameworks that combine textual information with sensor, process or environmental data.

2.1. Data Sources and Search Strategy

The primary data source was the Web of Science Core Collection. Searches were conducted over all WoS collections and covered the period from 2013 to October 2025. This time window was chosen because it encompasses both the consolidation of machine-learning and deep-learning approaches in safety-critical domains and, from approximately 2020 onwards, the emergence and rapid expansion of transformer-based models and LLMs, highlighting the recent acceleration of LLM/NLP applications.

The search strategy was iterative and was refined in several steps. In an initial phase, we used topic searches combining core OSH phrases with AI-related terms. Specifically, we ran separate queries in which the OSH term set “occupational risk prevention”, “occupational safety”, and “workplace safety”, was combined using Boolean AND with the AI term set “artificial intelligence”, “machine learning”, “deep learning”, “natural language processing”, “large language models”, and “LLM”.

We then restricted the AI term set to “natural language processing”, “large language models” and “LLM” and repeated the queries with “occupational safety” and “workplace safety”. This step captured the subset of studies in which NLP or LLMs are explicitly mentioned, enabling a detailed analysis of their data sources, model architecture and safety-related tasks.

Because OSH terminology varies across disciplines and sectors, we expanded the OSH term set beyond the expressions above. We ran an additional broad query with Topic (OSH term set): “occupational risk prevention” OR “workplace safety” OR “hazard identification” OR “incident analysis” OR “risk management” OR “EHS” OR “health and safety”; AND Topic (AI term set): “artificial intelligence” OR “machine learning” OR “deep learning” OR “natural language processing” OR “large language models” OR “LLM”.

This broad query retrieved 4,897 records and was used to quantify the magnitude and diversification of recent AI/ML/NLP work in OSH-related domains. The subset of records involving text-centric methods, LLMs or multi-source data integration relevant to occupational risk prevention was then examined in detail during screening.

We complemented the Web of Science search with targeted Google Scholar queries and backward/forward citation chasing. Reference lists and citation networks of key recent AI/NLP/LLM safety studies were used to identify additional publications that met the eligibility criteria. All records identified through these complementary routes were subjected to the same screening and selection process as database-retrieved records.

No formal language restrictions were imposed at the search stage; however, the vast majority of included studies were published in English.

2.2. Eligibility Criteria

We primarily considered peer-reviewed journal articles indexed in Web of Science. In line with a scoping-review approach, we also allowed the inclusion of high-quality conference papers, book chapters and technical reports when they (i) introduced novel AI/NLP/LLM methods, data sets or pipelines, or (ii) provided industrial case studies directly relevant to occupational risk prevention.

Records published between 2013, and October 2025 were eligible. This period captures the progressive adoption of machine-learning and deep-learning methods in safety analytics and the more recent use of transformers and LLMs in OSH-relevant tasks.

Studies had to address occupational or process safety broadly understood. Works focused purely on clinical patient safety, general medical decision-making or other non-occupational risk domains were excluded unless an explicit occupational or workplace context was present.

Eligible studies had to employ at least one AI-related method, such as machine learning, deep learning, NLP, LLMs, topic modelling, embedding-based retrieval, computer vision, or vision–language models, to analyze or model safety-related data. Given the primary focus of this review, priority was given to studies where: (i) unstructured safety narratives or text fields (e.g., accident reports, incident descriptions, investigation narratives, job descriptions, safety observations) were the main data source; or (ii) textual data formed an explicit component of a multi-source or multimodal risk-assessment framework (e.g., combined with sensor, process, or environmental data).

A small number of influential AI-based safety reviews and multi-source frameworks without a dominant textual component were retained when they provided essential methodological context for the development and deployment of NLP and LLM approaches in occupational risk prevention.

To be included, studies had to provide sufficient methodological and contextual detail to support analysis of: data sources, AI/NLP/LLM methods used, safety-related tasks and outcomes, and at least a qualitative assessment of model performance, advantages and limitations. Opinion pieces, editorials, non-peer-reviewed summaries, purely conceptual papers without empirical or methodological content, and works relying solely on basic descriptive statistics or traditional regression models (without AI, ML, DL, NLP or LLM components) were excluded.

Applying these criteria ensured that the final corpus represented the diversity of AI-driven, and particularly text-centric, approaches to occupational risk prevention while maintaining a clear OSH focus.

2.3. Screening and Selection Process

All records retrieved were deduplicated. The subsequent screening and selection proceeded in two stages. Titles and abstracts were screened against the eligibility criteria to remove clearly irrelevant items. At this stage, we excluded, for example, clinical or patient-safety studies without occupational context, generic AI or computer-science papers with no explicit link to OSH or process safety, and articles where the analytical methods were limited to classical statistics or non-AI approaches.

The full texts of the remaining articles were examined to verify that an AI/ML/DL/NLP/LLM or vision/vision–language model was applied; the data analyzed were occupational, process-safety or workplace-related (e.g., accident or incident reports, near-miss databases, hazard observations, FRCP/PSIF datasets, occupational exposure or health records, job postings, safety-management documents, or multi-source safety data); and the study provided enough methodological and contextual information to allow extraction of sector, data type, model family, safety-related task, and main performance and implementation insights.

After duplication and full-text assessment, 126 primary studies met all eligibility criteria and were included in the review.

2.4. Data Extraction and Synthesis

A structured data-extraction template was developed to ensure consistency across studies. For each of the 126 records included, we extracted:

• Bibliographic and contextual information, such as first author, year of publication, country or region (when reported), publication venue and study type (methodological paper, empirical case study, review, or framework proposal), as well as industrial sector and setting: dominant sector(s) addressed (e.g., aviation, construction, mining, chemical and process industries, manufacturing, transportation, healthcare, public sector and other services) and any specific workplace or process characteristics relevant to OSH.
• Data sources and modalities, including type and origin of the safety-related data (e.g., free-text accident/incident reports, near-miss and hazard observations, PSIF/FRCP datasets, occupational injury or disease registries, compensation claims, exposure or environmental monitoring data, job descriptions, safety-inspection reports, training materials, video or image data for PPE/unsafe-condition detection, sensor and process data in multi-source frameworks). Particular attention was given to whether unstructured text was the primary data source or part of a multimodal pipeline.
• AI/NLP/LLM and vision methods: main model families employed (e.g., classical supervised machine learning; topic models and other unsupervised text-mining techniques; word and sentence embeddings; recurrent or convolutional neural networks; transformer-based NLP models; LLMs and retrieval-augmented generation pipelines; ensemble models; computer vision and vision–language architectures) and any domain adaptation, fine-tuning or strategies reported.
• Analytical objectives (e.g., automated classification of incident types or causes; extraction of causal chains and contributing factors; topic modelling of safety concerns; prediction of accident occurrence, likelihood or severity, including PSIF and FRCP levels; risk-index estimation; early warning and anomaly detection; generation of job- or task-specific safety guidance and reports; PPE compliance or unsafe-condition detection; monitoring of safety-culture or safety-climate indicators).
• Evaluation and performance metrics (e.g., accuracy, precision, recall, F1-score, AUC, confusion matrices) and/or qualitative assessments (e.g., expert validation, comparative analyses with baseline methods, user or practitioner feedback) used to evaluate model performance and practical utility.
• Advantages, limitations and implementation aspects, such as improved prediction accuracy, better handling of unstructured narratives, ability to integrate multi-source data, explainability and data quality, under-reporting, representativeness issues, transparency and explainability challenges.

Following extraction, studies were first grouped by industrial sector to reflect the domain-specific context in which AI, NLP and LLMs are currently being deployed for occupational risk prevention. Within each sector, we then organized the literature by methodological family (text-mining and ML, deep-learning NLP, multi-source data-integration frameworks, LLM-based and RAG-based systems, and vision/vision–language approaches). Thus, the structure of the Results section allows both: (i) sector-specific narratives that highlight how AI-driven text analytics and LLMs are being tailored to particular hazards, data infrastructures and regulatory environments; and (ii) a cross-cutting synthesis of methodological trends, gaps and challenges, and real-world implementation in safety-critical contexts.

3. Results

As pointed out by Zhao et al (2018), the benefits of using NLP methods in occupational risks prevention are multiple; for example, these approaches allow valuable information to be extracted and processed from large amounts of data. Future research directions include pattern recognition, in-situ identification of actual events, and fully automated methods (Zhao et al., 2018). More specifically, previous reviews have highlighted the high potential of AI, LLMs and NLP methods in different areas of occupational risk prevention, as summarized in Table 1, such as exploring the impact of NLP applications in the field of aviation safety (Yang & Huang, 2023; Kierszbaum & Lapasset, 2020), and in other safety-critical industries such as transport, medical and construction (Ricketts et al., 2023), as well as for occupational injury analysis (Khairuddin et al., 2022), unveiling the influential aspects of this field through descriptive and scient metric analyses (Sarkar & Maiti, 2020).

LLM and AI are demonstrating great potential for development in different areas of occupational risk prevention, in sectors such as aviation and construction, even in specific risks such as Fall from Height (FFH), and in the chemical industry, as well as in the transport system, including railway, in the nuclear power generation sector, and for the protection of mine workers and to avoid medical errors.

Among the advantages that we can find in the use of methodologies based on AI, LLM and NLP in the field of occupational risk prevention, some are particularly interesting due to their high general applicability to multiple sectors, such as the generation of risk maps in real-time. For example, the application of dynamic real-time analysis using multimodal data fusion to enhance occupational risks prevention through the development of risk maps for workplaces, using machine/deep learning techniques by analysing data from diverse sources such as images, videos, documents, mobile applications and sensors/IoT. Thus, the combination of computer vision, NLP techniques, and sensor data analysis enables automated root cause identification, damage prevention, and disaster recovery, dynamically updating risk assessments in real-time. (Dalal & Bassu, 2020).

It is also worth mentioning that an important part of the success of the application of LLM and NLP-based methods lies in their ability to extract and analyze information in an automated way from large datasets contained in reports (e.g., accident reports), where the information can be structured to address a variety of problems, such as the limitations of generic and static checklists, which often do not apply to specific workplace contexts (Westhoven & Jadid, 2023), or mor interestingly, the information may not have been previously structured.

Thus, in relation to the use of unstructured information, recent research highlights innovative integrations of AI, specifically through NLP and Machine Learning (ML), to refine safety and risk assessments. For example, Kamil et al. (2023) combine a variety of NLP and text mining techniques with fuzzy set theory to transform unstructured accident reports into useful data, a methodology that contrasts with others used by Hou et al. (2022), who rationalize incident classification using NLP techniques for text vectorization. On the other hand, Paraskevopoulos et al. (2022) extend the functionality of AI in safety management by introducing a multimodal architecture that synergizes textual and visual data, distinguishing it from other studies primarily focused on text. In addition, Zhao et al. (2020) and Macedo et al. (2023) extend text analysis in different ways, as Zhao focuses on summarizing accident reports, while Macedo aims to correct inaccuracies in report. Furthermore, Baker et al. (2020a) and G. Liu et al. (2021) both refine data extraction and prediction methods, but differ in their approaches, as Baker emphasizes predictive modelling for safety outcomes, while Liu explores causal relationships using clustering techniques. Similarly, Dorsey et al. (2020) and Ekramipooya et al. (2023) aim to improve data quality and analysis efficiency through the use of NLP and AI methods.

Next, a perspective is presented on the impact of methodologies based on AI, LLM and NLP on the advancement of occupational risk prevention in different industrial sectors is presented.

Beyond sector-specific applications, the recent literature reveals several cross-cutting methodological patterns in how AI, NLP and LLMs are being leveraged for occupational risk prevention. First, there is a clear movement from shallow text-mining and keyword-based approaches towards deep contextual representations and transformer-based models for accident and incident narratives. For example, Song et al. (2024) use the KoBERT model to classify occurrence types from Korean industrial accident cases and explicitly link these classes to prevention plans, while Khairuddin et al. (2024) develop an optimized deep-learning prediction model that contextualizes injury severity from occupational accident reports. In parallel, Kim et al. (2024) and Li et al. (2025) demonstrate how transformer-based language models can be used for occupation and job-code classification using working-conditions surveys and job postings, thereby enriching OSH risk assessments with standardized occupational information. At the risk-modeling level, ensemble-learning frameworks and multi-source data integration are becoming increasingly prominent, as illustrated by Sarker et al. (2025) in the context of potential accident severity and FRCP classification, by Kamil et al. (2024) in multi-source incident-likelihood analysis, and by Lu et al. (2024) in data-driven coal-mine environmental safety risk assessment systems.

Second, there is an emergent family of hybrid models that explicitly combine AI-based text analytics with established risk-assessment or safety-engineering frameworks. Recent examples include the use of improved LDA topic models coupled with Bayesian networks to identify and propagate chemical safety risk factors (Zhou et al., 2025), Analytical Hierarchy Process (AHP) studies that quantify the relative contribution of human error, environmental conditions and equipment failure to mine accidents (Bekal et al., 2024), and frameworks for process risk assessment that integrate prior hazard information into chunk-based text-mining models (Sahoo et al., 2024). At a more advanced level, LLMs are increasingly embedded within accident-investigation and safety-management workflows, as shown by Du and Chen (2025) in coal-mine accident risk analysis, Ni et al. (2024) in the development of Industrial-Internet-based process-accident indicators, Liu et al. (2025) in HFACS-guided Chain-of-Thought accident investigation for general aviation, and by Baek et al. (2025) and Bernardi et al. (2025) in retrieval-augmented LLM frameworks for construction safety guidance and job safety report generation. Complementing these text-centric approaches, vision–language models that detect PPE compliance and unsafe behaviors from images and video (Chen et al., 2024) illustrate how language models can be coupled with computer vision to provide a multimodal foundation for proactive occupational risk prevention.

3.1. Aviation

In the field of aviation safety, Miyamoto et al. (2022) and Dong et al. (2021) both used NLP techniques to analyze aviation safety reports. Miyamoto et al. focused on categorizing the causes of flight delay using clustering techniques, revealing maintenance issues as the primary cause. In contrast, Dong et al. combined NLP with deep learning models to automate the identification of primary factors in incident reports, demonstrating superior performance over traditional methods but limiting their scope to the most frequent incident categories. Moreover, Jiao et al. (2022) introduced a novel classification scheme using the XGBoost classifier and OC-POS vectorization to identify risk factors from Chinese aviation reports, indicating great potential for broader applications. Similarly, Kierszbaum et al. (2022) developed a compact, domain-specific language model, demonstrating that specialized pre-training can effectively address the scarcity of domain-specific data in aviation safety NLU tasks, highlighting a trend towards creating more specialized NLP and AI tools tailored to specific data challenges in aviation safety. In addition, Madeira et al. (2021) investigated human factors in aviation incidents, using a hybrid approach of semi-supervised and supervised learning to tackle the challenge of limited labelled data sets, a common issue in AI applications in safety analysis. This study aligns with the work of Rose et al. (2020), who also used NLP and clustering to categorize and visualize safety narratives, but with a focus on integrating numerical and text-based data to enhance accident investigation processes. Liu et al. (2025) have significantly accelerated and improved the efficiency of general aviation accident investigations by integrating the HFACS framework into chain-of-thought prompts using large language models (LLMs). Their HFACS-CoT+ approach outperforms basic prompting strategies and, in some cases, human experts. Bernardi et al. propose a novel RAG-based architecture in their work that generates occupational safety reports from unstructured accident descriptions. By evaluating multiple large language model (LLM) families and incorporating models into the Aviation Safety Reporting System (ASRS) dataset, the study provides robust empirical evidence in support of using domain-specific AI solutions to improve accident analysis and decision-making. These studies highlight a significant trend towards using advanced AI and NLP methods to dissect and understand large volumes of aviation safety data, as it is shown in Table 2.

3.2. Construction

Additionally, occupational risk prevention (ORP) in the construction industry has a wide range of research that incorporates AI advances in safety management, moving towards automated, accurate and effective methods, as resumed in Table 3 and discussed below. Despite methodological diversity, the literature reveals converging trends, allowing research to be grouped into four key areas: text mining and ML, knowledge representation, multimodal AI, and large-scale language model (LLM) applications. Collectively, these studies demonstrate a shift toward automated methods in management, accurate and effective for risk identification and security management, although shortcomings also exist, such as inconsistencies in preprocessing workflows, the unlimited use of unsupervised NLP methods, and the underutilization of machine learning models (Shayboun et al., 2025).

Text-mining and machine learning techniques have been widely applied to classify accidents and extract risk factors from incident narratives. Early studies highlighted the potential of AI for information retrieval from construction documents (Fan & Li, 2013; Tian et al., 2023), while ensemble classifiers improved precision in identifying accident causes and safety risks (Zhang et al., 2019; Wang et al., 2021). Deep learning approaches further enhanced predictive accuracy and interpretability: Baker et al. (2020) employed Convolutional Neural Networks (CNNs) and Hierarchical Attention Networks to analyze accident reports, enabling visual interpretation of model predictions to identify injury precursors. Furthermore, Fang et al. (2020) and Gadekar & Bugalia (2023) improved text classification in construction safety reports, focusing on the use of Bidirectional Transformers (BERT) for deep learning-based text classification and innovating with a semi-supervised model, respectively, achieving high accuracy with reduced dependence on pre-labelled data. Advanced NLP preprocessing combined with novel AI techniques has also improved the effectiveness of construction safety analyses (Cheng et al., 2020). In the domain of metro construction, Xu et al. (2021) applied text mining with an information entropy–weighted term frequency metric to extract safety risk factors, providing a quantitative tool for large-scale risk assessment.

Structured knowledge representation using ontologies, knowledge graphs, and named-entity recognition (NER) has emerged as a powerful approach for automating safety management. Thompson et al. (2020) proposed a construction-specific NER scheme to structure free-text safety data into actionable strategies. Shen et al. (2022) introduced an innovative integration of Building Information Modeling (BIM) with an ontology-based safety rule library and NLP, creating a dynamic safety rule-checking system capable of automatically identifying hazards on construction sites. Graph-based approaches have also shown advantages over deep learning for certain risk domains; for example, Ben Abbes et al. (2022) used NLP and knowledge graphs to analyze the DBkWik database (40,000 wikis) for Fall From Height (FFH) risk, efficiently extracting critical safety information and addressing some limitations of deep learning methods. These approaches enable systematic structuring of heterogeneous safety information, supporting proactive hazard mitigation and compliance monitoring.

Visual and multimodal AI extends hazard detection beyond textual narratives. Zhong et al. (2023) developed a ResNet101–LSTM attention model to translate video sequences into natural language descriptions of on-site activities, allowing automated identification of unsafe behaviors. Such multimodal approaches complement text-based methods and are critical for monitoring complex or large-scale worksites where conventional reporting is insufficient.

LLMs are rapidly being applied to accident classification, causal pattern extraction, summarization, and safety training. GPT-based models have been used to classify accident types, uncover latent hazard structures, and analyze OSHA narratives (Salles et al., 2024; Smetana et al., 2024; Yhoo et al., 2024), while the AIR Agent automates extraction of accident categories from subway reports (Zhang et al., 2025). Retrieval-augmented LLMs can generate safety guidance and training materials of comparable or superior quality to expert-authored documentation (Uhm et al., 2024; Baek et al., 2025). Embedding techniques, such as SBERT, allow analysis of discrepancies between inspection reports and actual incidents (Elizabeth et al., 2025), and scenario-based LLM platforms provide validated training environments to strengthen safety decision-making (Naderi & Shojaei, 2025).

In addition to improving classification and information-retrieval tasks, LLMs are beginning to automate the generation of construction safety guidance. Baek et al. (2025) presents an automated safety risk management guidance framework that combines a retrieval module with a large language model in a Retrieval-Augmented Generation (RAG) architecture. Their system retrieves relevant reference documents from a large database of 64,740 construction accident cases and associated safety materials and then uses an LLM to generate tailored safety risk management guidance for specific work activities and equipment. By performing domain adaptation and instruction-tuning, the authors demonstrate that the LLM can generate guidance that is consistent with construction safety experts while significantly reducing the time required to prepare task-specific job-hazard analyses. This work illustrates how LLMs, when combined with robust retrieval mechanisms and domain-specific corpora, can move beyond passive text analysis to actively support the design of prevention measures and the dissemination of context-aware safety information in construction projects.

3.3. Chemical, Mines and Other High-Risk Industrial Environments

The integration of AI and NLP in chemical industry safety has the potential to enhance occupational and environmental safety (see Table 4). Thus, Kamil et al. (2023) used NLP, Interpretive Structural Model (ISM), and probabilistic techniques to predict and analyze fire and explosion risks, leveraging accident databases for predictive accuracy in safety management practices. On the other hand, Kabir et al. (2023) improved the accuracy of flare system failure analyses in the oil and gas industry by integrating traditional Fault Tree Analysis (FTA) with Dynamic Bayesian Networks (DBNs). In contrast, Kumari et al. (2022) advanced incident prediction by means of Artificial Neural Networks (ANNs) for cause and sub-cause analysis, surpassing traditional models to offer causation clarity. Moreover, B. Wang & Zhao (2022) introduced a novel deep learning framework combining BERT, BiLSTM-CRF, and CNN models to automate the extraction and classification of risk factors from accident reports in confined spaces, addressing the manual labor-intensive and subjective traditional analysis. Additionally, Xu et al. (2022) and Jing et al. (2022) utilized deep learning for analyzing accident causes, applying a CNN model to classify causes and deploying a combination of LSTM and attention mechanisms to enhance text classification of chemical accidents, respectively.

Furthermore, X. Luo et al. (2023) explored the use of NLP to automate the analysis of chemical accidents, categorizing risk factors to support decision making in risk analysis. Also, Macêdo et al. (2022) used BERT models for text mining to enhance quantitative risk analysis in oil refineries. Lastly, Song and Suh (2019) innovated in the detection of anomalies in accident reports by applying a text mining-based method to examine the narratives of accident reports.

Recent work on data-driven safety and occupational risk prevention in the process industries converges on the integration of heterogeneous data sources, advanced text mining and probabilistic modelling to improve prediction, assessment and control of accidents. Thus, Kamil et al. propose a Safety 4.0 framework that combines natural language processing of CSB loss-of-containment narratives with operational sensor data to build multi-source likelihood models, showing that inadequate written procedures and management failures are highly sensitive drivers of LOC events (Kamil et al., 2024). In coal mining, Lu et al. (2025) develop a dynamic environmental safety risk assessment system that fuses expert judgements, online monitoring and subjective reporting through fuzzy linguistic transformation, multi-criteria weighting and grey clustering, enabling real-time risk status updates and critical risk identification. At the plant level, Ni et al. (2024) leverage Industrial Internet infrastructures to operationalise major accident indicators and, using STAMP and a large language model to retrospectively analyse 212 accident reports, demonstrate SMART-compliant indicators that are empirically linked to accident patterns. Text mining of safety reports is further extended by Sahoo et al. (2024), who encode prior hazard knowledge in rule-based chunking to extract fault-related phrases and then use unsupervised and semi-supervised learning to reconstruct chains of events and fault trees with high agreement with expert HSE assessments.

In parallel, several contributions focus on mining large accidents and hazard datasets to support preventive decision-making. Song et al. (2024) show that KoBERT-based models can automatically classify occurrence types in Korean industrial accidents with high accuracy, reducing subjectivity and noise in national statistics and strengthening the basis for prevention planning. Zhou et al. (2025) apply an improved LDA topic model to chemical accident reports, identifying key risk factors and then using association rules and Bayesian networks to map their causal structure and critical paths, overcoming the subjectivity and limited scalability of traditional expert-based analyses. At mine level, Kar et al. (2024) use the Analytic Hierarchy Process on a decade of Indian mining accidents to quantify the relative contribution of human error, environmental conditions and equipment faults, finding human error to be the dominant factor across accident types and transport machinery to be the most critical alternative. Finally, Du and Chen (2025) integrate large language models, Apriori association rule mining and Bayesian networks on coal mine accident reports, extracting a rich hierarchy of risk factors and primary drivers linked to on-site safety management, procedure execution and supervision, and argue for policy responses centred on enforcement, training and data-driven early-warning systems. Together, these studies illustrate a rapid shift from purely retrospective, expert-driven investigation towards AI-enabled, multi-source and probabilistic frameworks that support proactive, system-level occupational risk prevention.

3.4. Transport System

The application of NLP and AI has the potential to enhance the accuracy and efficiency of risk assessment and safety management in Transport Systems (see Table 5). Thus, Hughes et al. (2019) used an AI-based model to extract and categorize terms from multilingual incident reports through the application of NLP techniques, achieving a high accuracy rate in categorizing safety incidents in public transport. Also, Valcamonico et al. (2022) also enhanced road safety analysis by integrating Hierarchical Dirichlet Processes and Doc2Vec with machine learning classifiers, showing how combined models can better balance accuracy and explainability in automated report classification. Moreover, Jidkov et al. (2020) focused on maritime risk assessment, employing deep learning and various NLP techniques to capture, process, and analyze data related to maritime safety events such as piracy, hijackings, and smuggling, improving incident classification and information extraction. Meanwhile, Wang & Yin (2020) employed text mining and automatic association rules such as the FP-Growth algorithm to uncover key risk factors in China’s transport sector, providing insights into systemic issues affecting safety. Additionally, Zhang et al. (2021) introduced the use of NLP and deep learning to analyze aviation accident reports with predictive purposes and safety management in aviation. More recently, Ricketts et al. (2022) proposed the use of NLP, rule-based phrase matching and a trained NER model to enhance hazard identification in HAZOP studies of aircraft subsystems, approaching the continuous model refinement and more efficient safety actions.

Specifically, in relation to railway safety and risk prevention, NLP and AI techniques have recently been used to innovate in incident prediction and management. For example, Hughes et al. (2018) developed a semi-automated classification system for close call reports in the GB railway industry, using NLP to associate incident reports with bow-tie accident causation models, with practical applications in categorizing a vast array of unstructured safety-related text. In contrast, Figueres-Esteban et al. (2016) used visual text analysis to extract safety information from the GB railways' Close Call System, highlighting its potential to identify risks despite the linguistic variation different reporter groups. Also, Wu et al. (2020) introduced NLP methods to improve subway accident decision-making processes in metro accidents with high precision in retrieving relevant past cases and advancing automated accident response systems. Moreover, Heidarysafa et al. (2018) applied deep learning to enhance the accuracy of accident labelling in the US railway sector and advanced the automatic classification of accident causes from narrative texts. Also, Ebrahimi et al. (2023) used NLP and Random Forest to develop a machine learning model capable of predicting evacuation needs following hazardous materials incidents on railways, mapping causal evacuation factors to improve emergency management. Furthermore, Hua et al. (2019) and Liu & Yang (2022) used text mining to improve risk identification in railway safety, extracting accident risk factors from Chinese railway accident reports through convolutional neural networks, and using deep learning techniques to quantify risk relationships in British railway incidents, respectively.

Kim's study (2023) used textual network analysis to examine the main issues related to death from overwork reported in the Korean media over a 10-year period in the Big Kinds database. Four themes were identified through theme modelling using the NetMiner 4 programme. The results revealed that postal workers, civil servants and delivery drivers are particularly susceptible to death from overwork.

3.5. Healthcare and Assistive Services Systems

The integration of NLP and AI techniques in healthcare has demonstrated substantial potential for preventing medical errors, enhancing patient safety, and supporting occupational health in clinical environments (see Table 6). For example, Cohan et al. (2017) employed convolutional and recurrent neural networks with an attention mechanism to analyze complex clinicians’ narratives, effectively identifying and categorizing harmful events. This approach not only improved error detection in large datasets but also facilitated root cause analysis and resource allocation, thereby contributing to the prevention of patient harm.

Similarly, Denecke (2016) highlighted the utility of NLP for processing critical incident reports, which are often underutilized due to the time-consuming and complex nature of manual review. By mapping incident reports to the International Classification of Patient Safety (ICPS) and employing text mining techniques, the study enabled semantic annotation, faceted search, and automated event detection, thereby enhancing both patient safety and quality of care.

Recent advances have further improved anomaly detection in electronic health records (EHRs), enhancing both patient safety and data reliability. Niu et al. (2024) developed EHR-BERT, outperformed existing models by reducing false positives, improving detection accuracy, and minimizing information loss, demonstrating the value of advanced NLP models in safeguarding patient care.

Beyond patient-centered applications, NLP and AI have also contributed to predicting occupational health risks. Sen et al. (2024), for instance, developed ERG-AI, an AI/ML pipeline combining multi-sensor posture data, uncertainty estimation, and large language model-generated recommendations to predict long-term worker postures and communicate associated risks. Evaluated on the DigitalWorker Goldicare dataset (114 workers, 2913 hours), ERG-AI delivered accurate, uncertainty-aware predictions while maintaining low energy consumption, providing personalized and interpretable health recommendations.

Finally, the adoption and perception of AI tools among clinicians have been systematically assessed. Egli et al. (2025) conducted an anonymous survey of Swiss healthcare professionals, revealing that 32.8% reported frequent use of large language models (LLMs), particularly among younger, male, and research-active clinicians. The study identified administrative and analytical support as primary benefits, while ethical considerations and output reliability emerged as key challenges.

Taken together, these studies underscore the transformative role of NLP and AI in healthcare and assistive services, from improving patient safety and clinical decision-making to enhancing occupational health, while highlighting the importance of user engagement, transparency, and ethical considerations in AI deployment.

3.6. Other Sectors

As shown in Table 7, other sectors, such as nuclear energy and mining, also benefit from the integration of AI-based methods, particularly LLM and NLP. In relation to the application of NLP techniques to the nuclear power generation sector, Zhao et al. (2019; 2018) advanced the field by integrating NLP and multimodal data fusion to automatically identify causal relationships in event reports. They used arule-based expert system, the Causal Relationship Identification (CaRI), to effectively capture causal associations with a success rate of 86%. On the other hand, Dalal & Bassu (2020) explored the development of "risk maps" by applying machine learning models to analyze data from sensors and computer vision systems to achieve a dynamic real-time capability to identify risks and prevent workplace accidents. The combination of NLP and AI methods in the field of occupational risk prevention has also recently led to several studies related to mine safety. Thus, Ganguli et al. (2021) carried out automatic data analysis from Mine Health and Safety Management Systems (HSMS) using NLP and Machine Learning (ML), specifically through the development of nine Random Forest (RF) models, demonstrating high accuracy and improved incident categorization.

Recent work has also explored LLM-based decision support in more heterogeneous industrial environments. Bernardi et al. (2025) propose a Hum-AI/Gen-AI framework in which an LLM is combined with retrieval and explanation modules to generate job safety reports that summarize hazards, recommend preventive measures and document safety-critical activities. Their approach, implemented in an Information Systems Frontiers case study, retrieves relevant regulations, standards and prior incidents, and prompts the LLM to synthesize concise, context-specific safety recommendations while exposing the underlying evidence used to generate each suggestion. In the aviation domain, Liu et al. (2025) develop an HFACS-guided Chain-of-Thought (CoT) accident-investigation framework in which LLMs reason step-by-step through witness narratives and investigation texts to allocate causal factors to HFACS categories. The authors show that structuring prompts according to HFACS levels and requiring explicit CoT explanations substantially improves both accuracy and interpretability relative to direct-answer prompting, providing a promising template for LLM-assisted accident investigation in other high-hazard sectors.

In contrast, Shekhar and Agarwal (2021) applied text mining of fatality reports to enhance safety in Indian mines, identifying trends and patterns and highlighting the most vulnerable worker demographics and high-risk times periods. Furthermore, Qiu et al. (2021) combined text mining with complex network analysis to identify and quantify factors contributing to coal mine accidents, revealing complex interaction mechanisms and critical causal links, and providing a detailed map of accident causation pathways.

3.7. Severity, PSIF and proactive risk prediction

A growing body of work no longer treats AI and NLP as purely descriptive tools for analyzing past accidents, but rather as instruments for proactively estimating accident severity, identifying potential serious injuries and fatalities (PSIF), and supporting fatality-risk control programs. Khairuddin et al. (2024) introduce an optimized deep-learning prediction model that contextualizes injury severity based on free-text occupational accident reports. By combining deep neural networks with carefully engineered features derived from accident narratives, their model achieves improved performance over conventional classifiers in predicting severity classes, thereby enabling safety practitioners to prioritize high-risk cases for investigation and control. Complementing this case-based approach, Parikh et al. (2024) propose an automatic PSIF identification framework that flags incidents involving potential serious injuries and fatalities in large incident databases. Their study shows that focusing on PSIF-related incidents captures underlying exposure to fatal hazards more effectively than relying solely on historical fatality counts, and that automated PSIF classification can support a shift from reactive to proactive safety management.

At the level of structured safety programs, Sarker et al. (2025) develop an ensemble-learning framework that integrates NLP-derived features from accident narratives with structured safety data to predict potential accident severity and assign accidents to Fatality Risk Control Program (FRCP) levels. The authors combine multiple classifiers and assess feature importance via Leave-One-Covariate-Out (LOCO) analysis, demonstrating that narrative-based features and FRCP-specific indicators significantly improve predictive performance. Their results highlight how NLP-enhanced models can support FRCP implementation by providing early warnings and helping safety managers focus on high-impact events. Related ensemble and multi-source approaches also appear in process-safety contexts, where heterogeneous data (incident narratives, process parameters, environmental indicators) are combined to estimate incident likelihood and severity (Kamil et al., 2024; Lu et al., 2024; Sahoo et al., 2024).

Crucially, these severity-oriented models are emerging across multiple industrial domains. In steel manufacturing, Sarker et al. (2025) show that integrating narrative features with FRCP categories can anticipate severe events before they occur, whereas in mining, Du and Chen (2025) combine LLM-based extraction of causal factors from accident descriptions with Bayesian networks to estimate the probability of severe coal-mine accidents under different control measures. In the broader manufacturing context, Song et al. (2024) demonstrate that transformer-based occurrence-type classification supports the design of targeted prevention plans, implicitly influencing severity distribution by reducing the frequency of hazardous occurrence types. Taken together, these contributions illustrate an important shift towards AI- and NLP-enabled models that aim not only to understand past accidents, but also to predict their potential severity and embed these predictions into structured safety programs such as PSIF monitoring and FRCP implementation.

3.8. Text Mining, Topic Modelling and Hybrid Risk-Assessment Frameworks

Recent studies highlight the value of combining advanced text-mining and topic-modelling techniques with established risk-assessment and decision-analytic frameworks. Zhou et al. (2025) propose an improved Latent Dirichlet Allocation (LDA) topic model tailored to chemical safety incident data, which is subsequently coupled with a Bayesian network to quantify the probabilistic relationships between latent risk themes and observable accident outcomes. By incorporating domain knowledge into the topic modeling and using the inferred topics as nodes in the Bayesian network, their framework supports both identification of critical risk factors and scenario-based reasoning about the effect of preventive measures. Similarly, Du and Chen (2025) use LLMs to extract causal factors, unsafe conditions and contextual information from coal-mine accident reports, which are then encoded as nodes and conditional probabilities in a Bayesian network. This hybrid LLM–Bayesian approach enables coal-mine safety practitioners to perform “what-if” analyses and to assess the impact of different control strategies on accident likelihood and severity.

In process-safety and chemical-engineering settings, Kamil et al. (2024) introduce a multi-source heterogeneous data integration framework for incident likelihood analysis that combines NLP features from incident narratives with process-operation and environmental data. Their approach leverages representation learning to map heterogeneous inputs into a unified feature space, and then applies machine-learning models to estimate incident occurrence probabilities under varying operating conditions. Sahoo et al. (2024) similarly demonstrate how prior hazard information can be encoded into chunk-based text-mining models for process risk assessment, where domain-specific hazard concepts guide the segmentation and representation of textual data. These works exemplify how text mining can be systematically aligned with process-safety knowledge to produce risk indicators that are both data-driven and interpretable.

In the mining sector, Bekal et al. (2024) use the Analytical Hierarchy Process (AHP) to quantify the relative contribution of human error, environmental factors and equipment failure to mine accidents in India. Although their study does not rely on deep learning, it illustrates how structured expert judgments and hierarchical modeling can complement data-driven text mining by providing formal weights for different risk categories. Lu et al. (2024) extend the hybrid paradigm further by constructing a data-driven coal-mine environmental safety risk assessment system that integrates multi-source heterogeneous data, including environmental sensor readings and operational factors, into an objective, dynamic and real-time risk index. When viewed together with Zhou et al. (2025), Du and Chen (2025), Kamil et al. (2024) and Sahoo et al. (2024), these studies show how NLP and topic modelling can be tightly coupled with Bayesian networks, AHP and other decision-analytic tools to operationalize complex risk assessment in chemical and mining industries.

Finally, these hybrid frameworks are beginning to interact with LLM-based reasoning in more sophisticated ways. Ni et al. (2024) exploit Industrial-Internet data and retrospective accident analysis to develop major process-accident indicators, using LLMs to assist in the categorization and interpretation of accident factors under STAMP-inspired structures and SMART criteria. In a similar spirit, Bernardi et al. (2025) embed LLMs within a RAG pipeline that retrieves relevant regulations, safety guidelines and historical cases to generate job safety reports and risk-mitigation suggestions, while explicitly logging the retrieved evidence to maintain transparency. Taken together, these contributions suggest an emerging paradigm in which text mining, topic modelling, Bayesian reasoning and LLM-based generation are combined to support interpretable and context-aware risk assessment.

3.9. Vision and Vision–Language Models for PPE Compliance and Unsafe conditions

While most of the studies in this review operate on textual data, recent advances in multimodal AI show that integrating vision and language models can greatly enhance the detection of unsafe behaviors and conditions. Chen et al. (2024) propose a vision–language model for interpretable and fine-grained detection of safety compliance in diverse workplaces. Their model leverages a CLIP-style architecture to perform zero-shot detection of personal protective equipment (PPE) items and unsafe configurations by aligning visual features with text prompts describing compliant and non-compliant conditions. The authors demonstrate that the model can accurately identify missing PPE, improper usage and unsafe postures across different industrial settings without requiring extensive task-specific training data, while also providing textual rationales that explain why a given frame is flagged as compliant or not. By explicitly encoding safety concepts in natural language, this approach enables transparency and human-interpretable feedback on vision-based compliance assessments.

These developments naturally extend the text-analytic approaches reviewed in earlier sections. For instance, multimodal frameworks could link vision-based PPE-compliance observations with incident narratives and near-miss reports, enabling models to correlate observed unsafe behaviors with subsequent accidents and thereby refine proactive risk indicators. In addition, occupation- and job-task-classification models trained on working-condition surveys and job postings (Kim et al., 2024; Li et al., 2025) could be combined with vision–language PPE detectors to tailor compliance criteria and risk thresholds to specific occupations, tasks and work environments. Such integration would support adaptive, context-aware safety monitoring in which camera-based systems not only detect missing PPE but also understand which PPE is required for a given job and why its absence increases risk. Although practical deployment raises important privacy, ethical and regulatory questions (see Section 4), vision–language models hold substantial promise for scaling proactive detection of unsafe conditions in complex workplaces.

4. Conclusions

This scoping review shows that AI, NLP and LLMs are reshaping occupational risk prevention across a wide range of industrial sectors. Early applications of text mining and traditional machine learning have been complemented by more recent transformer-based and ensemble-learning approaches that operate directly on unstructured OSH data such as accident narratives, inspection reports, work-condition surveys and safety guidelines. Emerging models predict accident severity, PSIF status and FRCP levels (Parikh et al., 2024; Khairuddin et al., 2024; Sarker et al., 2025), integrate heterogeneous sources of process and environmental data for incident-likelihood analysis (Kamil et al., 2024; Lu et al., 2024; Sahoo et al., 2024), and embed LLMs and Bayesian networks within accident-investigation and risk-assessment pipelines (Zhou et al., 2025; Du & Chen, 2025; Ni et al., 2024). Parallel developments in occupation and occurrence-type classification (Song et al., 2024; Kim et al., 2024; Li et al., 2025) and in vision–language PPE-compliance monitoring (Chen et al., 2024) expand the scope of AI-driven safety analytics beyond post-hoc analysis to encompass proactive monitoring and worker-centered prevention strategies.

Despite these advances, several limitations and risks associated with the use of AI, NLP and LLMs in occupational risk prevention must be acknowledged. Many models are trained on historical accident reports or incident databases that reflect under-reporting, incomplete causal information and sector- or country-specific biases. As a result, predictions of severity, PSIF, FRCP level or incident likelihood may inherit and amplify these biases, particularly for under-represented worker groups, subcontractors or informal sectors. Ensemble-learning and multi-source frameworks such as those proposed by Kamil et al. (2024), Lu et al. (2024), Sahoo et al. (2024) and Sarker et al. (2025) partly mitigate these issues by integrating diverse data sources, but they still rely on the quality and representativeness of the underlying data. Similarly, topic-model–Bayesian-network approaches (Zhou et al., 2025; Du & Chen, 2025) and Industrial-Internet-based indicator systems (Ni et al., 2024) may be sensitive to modeling assumptions, discretization choices and expert-defined structures.

LLM-based systems introduce additional concerns related to opacity, hallucinated content, and alignment with domain regulations and ethical principles. In RAG-enhanced safety-guidance and job-safety-report frameworks (Baek et al., 2025; Bernardi et al., 2025), the quality of the retrieval step and the coverage of the underlying document repository critically determine whether the generated guidance is accurate and compliant with current legislation. HFACS-guided Chain-of-Thought prompting (Liu et al., 2025) and vision–language PPE detectors (Chen et al., 2024) improve interpretability by exposing intermediate reasoning or by aligning decisions with human-readable safety concepts, but they still depend on careful prompt engineering, safety-specific evaluation and continuous monitoring to prevent erroneous or unsafe recommendations. From a regulatory perspective, the deployment of such systems in real workplaces must comply with data-protection and surveillance regulations, worker-participation requirements, and emerging AI governance frameworks. This underscores the need for robust human-in-the-loop designs in which OSH professionals critically review AI outputs, validate them against established risk-assessment methods (e.g. FRCP, PSIF, AHP-based ranking) and retain ultimate responsibility for safety-critical decisions.

Recent developments show that deep learning models applied to accident narratives can accurately predict injury severity categories and near-miss potential, going beyond traditional coded data and rule-based systems. In particular, NLP-based classifiers have been used to infer detailed severity labels, PSIF events and FRCP-type control-strength scores directly from free-text descriptions, supporting the prioritization of investigations and reinforcement of critical controls. These approaches complement other narrative-driven models that detect injury precursors and evacuation decisions from textual and multi-modal safety data, thereby turning unstructured information into actionable leading indicators. However, most studies are still based on single-organization datasets with class imbalance and limited external validation, underscoring the need for cross-industry benchmarks, transparent performance reporting and closer collaboration with practitioners to calibrate decision thresholds and embed these predictions within existing risk control protocols.

In the process and chemical industries, recent works show that text mining and deep contextual language models (e.g., BERT-type architectures) substantially improve the extraction of causal and contextual risk factors from accident and near-miss reports, outperforming traditional bag-of-words approaches and enabling the discovery of latent hazard themes through topic modelling. In parallel, coal-mine safety studies illustrate how multi-source data, structured accident causation frameworks and multi-level indicator systems, often weighted through expert judgement and data-driven methods and combined with AHP or human-factor models, allow more nuanced quantification of environmental and human-error-related risks. Beyond mining, hybrid risk-prediction frameworks are emerging that fuse indicators derived from text (extracted risk factors) with process, equipment and environmental data using probabilistic or mixture models to anticipate accidents proactively. Overall, these contributions point to hybrid NLP plus probabilistic/multi-criteria approaches as a key future direction for occupational risk prevention, while also highlighting that current solutions remain largely at prototype stage and require tighter integration with operational safety management, including explicit treatment of uncertainty and sensitivity analysis.

Recent LLM- and RAG-based systems in OSH move beyond analysing incidents to generating concrete preventive guidance. By combining domain-adapted embedding models for retrieving similar accident and job-safety cases with generative LLMs, these tools can automatically produce job hazard analyses, structured safety reports and checklists tailored to specific tasks, and even highlight likely root causes. They are also being piloted in conversational formats (e.g., chatbots) to support workers directly. Together, these developments signal a shift towards prescriptive AI in occupational safety, while underscoring the need for transparent pipeline descriptions (retrieval, prompting, generation, human-in-the-loop validation), rigorous control of hallucinations and domain adaptation, and the systematic use of curated knowledge bases and mandatory expert review before any recommendations are implemented.

Advances in workplace monitoring are moving from isolated IoT and vision systems towards contrastive vision–language architectures that jointly encode images and textual safety prompts. Models of this type can recognise multiple PPE items and unsafe behaviours simultaneously from surveillance images and, thanks to their prompt-based design, allow safety managers to formulate new monitoring queries (e.g., specific unsafe actions) without retraining. They complement traditional computer-vision approaches for PPE or posture detection, but also introduce important challenges around privacy and surveillance, potential bias in detection performance, false alarms that may undermine trust, and the need to integrate alerts into existing safety workflows under emerging ethical and regulatory constraints.

Current AI and LLM applications in occupational risk prevention still face important limitations related to data quality, representativeness and bias, since most models are trained on historical records from single organisations with under-reporting and uneven coverage of vulnerable workers. They also present challenges of transparency and explainability, domain adaptation and concept drift, together with the persistence of hallucinations and the risk of over-reliance on generative outputs, which makes robust governance, continuous monitoring and human-in-the-loop validation indispensable. In parallel, the growing regulatory and ethical scrutiny of general-purpose AI underscores the need for transparency, accountability and human oversight when deploying these tools in safety-critical contexts. Explicitly recognising these constraints offers a more realistic view of the current maturity of AI/LLM systems in OSH and supports their positioning as decision-support tools rather than autonomous safety authorities.

Despite these advances, several limitations and risks associated with the use of AI, NLP and LLMs in occupational risk prevention must be acknowledged. Many models are trained on historical accident reports or incident databases that reflect under-reporting, incomplete causal information and sector- or country-specific biases. As a result, predictions of severity, PSIF, FRCP level or incident likelihood may inherit and amplify these biases, particularly for under-represented worker groups, subcontractors or informal sectors. Ensemble-learning and multi-source frameworks such as those proposed by Kamil et al. (2024), Lu et al. (2024), Sahoo et al. (2024) and Sarker et al. (2025) partly mitigate these issues by integrating diverse data sources, but they still rely on the quality and representativeness of the underlying data. Similarly, topic-model–Bayesian-network approaches (Zhou et al., 2025; Du & Chen, 2025) and Industrial-Internet-based indicator systems (Ni et al., 2024) may be sensitive to modelling assumptions, discretization choices and expert-defined structures.

LLM-based systems introduce additional concerns related to opacity, hallucinated content, and alignment with domain regulations and ethical principles. In RAG-enhanced safety-guidance and job-safety-report frameworks (Baek et al., 2025; Bernardi et al., 2025), the quality of the retrieval step and the coverage of the underlying document repository critically determine whether the generated guidance is accurate and compliant with current legislation. HFACS-guided Chain-of-Thought prompting (Liu et al., 2025) and vision–language PPE detectors (Chen et al., 2024) improve interpretability by exposing intermediate reasoning or by aligning decisions with human-readable safety concepts, but they still depend on careful prompt engineering, safety-specific evaluation and continuous monitoring to prevent erroneous or unsafe recommendations. From a regulatory perspective, the deployment of such systems in real workplaces must comply with data-protection and surveillance regulations, worker-participation requirements, and emerging AI governance frameworks. This underscores the need for robust human-in-the-loop designs in which OSH professionals critically review AI outputs, validate them against established risk-assessment methods (e.g. FRCP, PSIF, AHP-based ranking) and retain ultimate responsibility for safety-critical decisions.

Future research should therefore prioritize (i) the creation of high-quality, representative and ethically sourced OSH datasets that cover diverse sectors, worker populations and countries; (ii) the development of hybrid frameworks that combine domain-knowledge-driven models (e.g. Bayesian networks, AHP, FRCP, PSIF) with AI-based text, vision and multimodal analytics; and (iii) systematic evaluation protocols that assess not only predictive performance but also fairness, robustness, interpretability and regulatory compliance. Particular attention should be paid to longitudinal validation of severity and PSIF prediction models, cross-sector generalization of LLM-assisted accident investigation methods, and integration of occupation-classification and working-conditions data into risk-assessment pipelines (Kim et al., 2024; Li et al., 2025). Finally, multidisciplinary collaboration between AI researchers, OSH professionals, regulators and worker representatives will be crucial to ensure that AI-enabled safety tools are designed, validated and deployed in ways that effectively reduce occupational accidents and diseases while respecting workers’ rights, autonomy and well-being.