A comparison of machine learning models for predicting Vibrio parahaemolyticus in oysters

Abstract

Vibrio parahaemolyticus, a major seafood pathogen, threatens public health as oyster consumption rises. We evaluated 14 machine learning models to predict its concentrations in oysters, achieving high accuracy (Concordance Correlation Coefficient, CCC > 0.85 training, > 0.9 testing, except bag-MARS) across diverse algorithms. Processing times varied from 23 min (KNN) to 162 min (bag-RPart), highlighting computational tradeoffs. Five top models-Elastic Net (EN), Random Forest (RF), XGBoost, Light Gradient-Boosting Machine (LGBM), and Cubist (39-92 min)-were selected for their performance and efficiency, forming a robust toolkit for shellfish safety monitoring. Variable importance and partial dependence plots identified sea surface temperature (SST) and wind as primary drivers, with SST thresholds of 16-26 degrees C driving proliferation and wind showing mixed effects (negative >4 m/s, positive >6 m/s). Precipitation, salinity (>19 ppm), and pH (7.5-7.7) played supplementary roles. Lagged variables (e.g., SST_imX_25) underscored temporal dynamics, supporting real-time monitoring and risk assessment strategies.