A data-driven predictive system for shellfish biotoxin risk assessment

Diarrhetic shellfish poisoning (DSP) toxins pose a major threat to the aquaculture industry and public health. Traditional monitoring methods suffer from delays and high costs, and intelligent early warning systems are urgently needed. The study utilised machine learning to predict shellfish biotoxin levels in shellfish by integrating multiple data sources, including phytoplankton monitoring, climatic variables, and historical toxin records. Data from seven shellfish production areas in Northern Ireland from 2011 to 2021, covering DSP-OA + dinophysistoxins and AZAs toxin concentrations, seven phytoplankton taxa or taxonomic groups, and three climatic variables was used to inform the model. Using the Lale Machine learning framework, the study constructed 42 initial modelling scenarios across regions, shellfish species, and toxin combinations. SHAP analyses quantified feature contributions and identified the main predictors across the retained scenarios. After data filtering and performance screening, model performance varied across the region-species combinations. The final regression models showed limited predictive performance, while the classification models showed stronger performance for the two evaluated region-species combinations, Belfast mussels and Dundrum mussels, with F1 scores of 0.766 and 0.950. The results of the study show predictive potential for toxin occurrence at the next sampling occasion, although further validation is required before operational early-warning application.