Predicting Paralytic Shellfish Poisoning on the West Coast of Canada

Abstract

Paralytic Shellfish Toxins (PSTs) are a major public health concern in contaminated shellfish. We developed and validated a predictive framework for PST risk in blue mussels (Mytilus edulis) along Canada’s west coast (2000–2020), comparing multiple machine learning and statistical models. Our study comprised three phases: (I) evaluating models with HPLC data at a 25 µg 100g⁻¹ detection threshold, (II) testing model transferability from legacy bioassay data to HPLC data, and (III) analyzing both total and individual toxin compounds using HPLC data at a 0 µg 100g⁻¹ threshold. Results showed cross-method predictions reduced model performance, while lower detection thresholds improved accuracy. Tree-based algorithms excelled with multivariate toxin data, whereas simple models performed best with univariate data. The ensemble model consistently matched the best individual model’s performance (AUC 0.942) across phases, serving as an effective automatic model selector despite varying optimal models.

Keywords: Paralytic Shellfish Toxin, Blue Mussels, Machine Learning, Time series