CABPT: Predicting Southern California Marine Biotoxins in Shellfish via Deep, Spatiotemporal, Multivariate Prediction

The threat of marine biotoxin contamination shellfish jeopardizes public health, ocean biodiversity, and aquaculture. This paper establishes the first California Marine Biotoxin Prediction Tool (CABPT), a novel and scalable AI-based framework for predicting the concentrations of domoic acid and saxitoxins in California. Development of the CABPT includes the following tasks: system designing, data sourcing and preprocessing, AI model building, and web server framework deploying. The CABPT integrates six data types, dissolved oxygen, pH, sea surface temperature, salinity, and precipitation, for predictive modeling in two Southern California regions, Santa Cruz and Santa Barbara. Categorical classification prediction was tested on two deep learning architectures (long short-term memory (LSTM), gated recurrent units (GRUs)) and two classical machine learning architectures (random forest (RF), gradient boosting (GB)). The results indicated deep learning models performed better than classical machine learning models. Specifically, LSTM models consistently outperformed all other models, particularly in handling class imbalance, achieving an average accuracy of 99.14% and performing more accurately on minority classes by an average of 50.92% compared to its deep learning counterpart, GRUs. Based on location-specific data from Santa Cruz and Santa Barbara and with time series cross-validation, the tool's 3–4-week forecast horizon aids in the timely and reliable flagging of shellfish species as dangerous or lethal for consumption. The CABPT has direct impact as a powerful architecture for marine safety and is poised to provide authorities and the public with actionable insights for proactive biotoxin management.