Machine learning-based modeling for monitoring and predicting the detection rate and severity of pathogenic...

Pathogenic microbial contamination in seafood presents persistent risks to food safety and public health. Conventional monitoring methods frequently lack adequate sensitivity and specificity, particularly under dynamic environmental conditions. To address these limitations, this study developed a machine learning (ML) framework utilizing microbial testing data from seafood samples collected in Yantai, China (2014-2024), with 2025 data reserved for external validation. The framework simultaneously predicts pathogen prevalence and severity levels. And, exploratory data analysis assessed variable distributions and statistical associations with detection rates. Time series analysis identified significant long-term trends and seasonal patterns, while spatial analysis revealed heterogeneity in contamination risk across administrative regions, particularly in coastal economic zones and major ports. Applying a dynamic threshold optimization strategy, detection rates were predicted using six ML algorithms: ROSE-LASSO, LightGBM, XGBoost, k-NN, CART, and SVM. LightGBM demonstrated optimal performance with 97.6% specificity and 99.9% positive predictive value (PPV) at an optimized threshold of 0.157. External validation confirmed its robustness, yielding 91.2% sensitivity, a 92.4% F1 score, and 93.6% PPV. For three-class severity classification, XGBoost achieved 93.7% overall accuracy and high sensitivity (80.0%) for high-risk samples, with an AUC of 0.989. By integrating spatiotemporal environmental features with adaptive thresholding, this framework provides a scalable approach for pathogen risk identification within coastal seafood supply chains, supporting early warning systems and evidence-based resource allocation.