Paralytic shellfish poisoning risk assessment in the west coast of Canada

Paralytic Shellfish Poisonings (PSPs) are a major public health concern requiring robust environmental monitoring. We developed and validated a machine learning framework to assess PSP risk in blue mussels (Mytilus edulis) along Canada's west coast. Our study comprised three experiments that tested the ability of 11 models to forecast risk using historical toxin data (2000-2020). Results showed that lower detection thresholds and the use of multivariate toxin profiles significantly improved predictive accuracy. Tree-based algorithms, in particular, excelled with this detailed data. A stacked ensemble model consistently matched the best individual model's performance, achieving an AUC (area under receiver operating characteristic curve) over 0.912 across all experiments and offering a robust solution for operational forecasting. Model interpretation revealed that recent toxin history and specific compounds like Neosaxitoxin (NEOSTX) and N-sulfocarbamoyl gonyautoxin-3 (C-2) were the most important predictors, aligning with regional ecological dynamics. This framework provides a powerful, data-driven tool for enhancing early warning capabilities and supporting proactive risk management.