Artificial Intelligence for Predicting and Prioritizing Microbiological Hazards in Food Safety Inspection Systems: A PRISMA-Compliant Systematic Review (2020–2025)

Background Foodborne diseases represent a persistent and escalating global public health challenge, with the World Health Organization estimating approximately 600 million cases and 420,000 deaths annually. Traditional food safety regulatory frameworks, characterized by reactive inspection models and conventional microbial testing, are increasingly inadequate in the face of complex international supply chains, emerging pathogens, and evolving consumer demands. Artificial intelligence (AI) and machine learning (ML) techniques offer a transformative paradigm for risk prediction and inspection prioritization, moving beyond the constraints of classical predictive microbiology to enable integration of heterogeneous, high-dimensional data sources for proactive hazard management. Objective This systematic review aimed to identify, critically appraise, and synthesize peer-reviewed evidence on the application and efficacy of AI/ML models for predicting microbiological hazard risks and optimizing food safety inspection prioritization within regulatory systems, published between January 2020 and December 2025. Methods A systematic search of Scopus, Web of Science Core Collection, and PubMed was conducted following PRISMA 2020 guidelines. Eligibility criteria were structured using the PECO (Population, Exposure, Comparator, Outcomes) framework. Data extraction encompassed study design, AI/ML architecture, data sources, validation strategy, and performance metrics. Risk of bias was assessed using the Prediction Model Risk of Bias Assessment Tool (PROBAST). Due to substantial clinical and methodological heterogeneity among included studies, a narrative synthesis was performed rather than a quantitative meta-analysis. Results Of 2,148 records identified, six studies met the inclusion criteria. Tree-based ensemble models (Gradient Boosted Decision Trees, Random Forest, XGBoost) consistently demonstrated strong discriminative performance for facility-level risk stratification with tabular surveillance data, achieving accuracy values of 0.69–0.85 and R² values exceeding 0.65. Deep learning architectures, particularly YOLOv5 for image-based detection, achieved approximately 94% accuracy on unstructured data. However, PROBAST assessment revealed that three of six studies (50%) carried high overall risk of bias, predominantly due to reliance on internal validation without external temporal or geographic validation, raising concerns about model generalizability and potential overfitting. Conclusions AI/ML models demonstrate substantial potential for enhancing microbiological risk prediction and inspection resource optimization in food safety. However, the current evidence base is limited by a pervasive lack of rigorous external validation, small study numbers, and heterogeneity in modeling approaches. For AI-driven systems to achieve trustworthy and scalable regulatory implementation, prioritizing robust validation standards, mandating explainable AI (XAI) frameworks for transparency and accountability, and investing in data governance infrastructure are essential prerequisites. Future research should focus on multi-site external validation, standardized benchmarking datasets, and real-world pilot deployments to bridge the gap between algorithmic promise and operational impact. Systematic Review Registration: No formal protocol registration was completed prior to review conduct. Methods were predefined in accordance with PRISMA 2020 and PROBAST standards. The authors recommend future updates register with PROSPERO.