Advancements in Artificial Intelligence and Machine Learning for Occupational Risk Prevention: A Systematic Review on...

Background: Occupational risk prevention is a critical discipline for ensuring safe working conditions and minimizing accidents and occupational diseases. With the rise of artificial intelligence (AI) and machine learning (ML), these approaches are increasingly utilized for predicting and preventing workplace hazards. This systematic review aims to identify, evaluate, and synthesize existing literature on the use of AI algorithms for detecting and predicting hazardous environments and occupational risks in the workplace, focusing on predictive modeling and prevention strategies. Methods: A systematic literature review was conducted following the PRISMA 2020 protocol, with minor adaptations to include conference proceedings and technical reports due to the topic's emerging and multidisciplinary nature. Searches were performed in IEEE Digital Library, PubMed, Scopus, and Web of Science, with the last search conducted on 1 August 2024. Only peer-reviewed articles published from 2019 onwards and written in English were included. Systematic literature reviews were explicitly excluded. The screening process involved duplicate removal (reducing 209 initial documents to 183 unique ones), a preliminary screening based on titles, abstracts, and keywords (further reducing to 92 articles), and a detailed full-text review. During the full-text review, study quality was assessed using six quality assessment (QA) questions, where articles receiving a total score below 4.5 or 0 in any QA question were excluded. This rigorous process resulted in the selection of 61 relevant articles for quantitative and qualitative analysis. Results: The analysis revealed a growing interest in the field, with a clear upward trend in publications from 2021 to 2023, and a continuation of growth into 2024. The most significant contributions originated from countries such as China, South Korea, and India. Applications primarily focused on high-risk sectors, notably construction, mining, and manufacturing. The most common approach involved the use of visual data captured by cameras, which constituted over 40% of the reviewed studies, processed using deep learning (DL) models, particularly Convolutional Neural Networks (CNNs) and You Only Look Once (YOLO). Conclusions: The study highlights current limitations, including an over-reliance on visual data (especially challenging in low-visibility environments) and a lack of methodological standardization for AI-based risk detection systems. Future research should emphasize the integration of multimodal data (visual, environmental, physiological) and the development of interpretable AI models (XAI) to enhance accuracy, transparency, and trust in hazard detection systems. Addressing long-term societal implications, such as privacy and potential worker displacement, necessitates transparent data policies and robust regulatory frameworks.