Random forest classification to determine environmental drivers and forecast paralytic shellfish toxins in Southeast Alaska with high temporal resolution.

Paralytic shellfish poison toxins (PSTs) produced by the dinoflagellate in the genus Alexandrium are a threat to human health and subsistence lifestyles in Southeast Alaska. It is important to understand the drivers of Alexandrium blooms to inform shellfish management and aquaculture, as well as to predict trends of PST in a changing climate. In this study, we aggregate environmental data sets from multiple agencies and tribal partners to model and predict concentrations of PSTs in Southeast Alaska from 2016 to 2019. We used daily PST concentrations interpolated from regularly sampled blue mussels (Mytilus trossulus) analyzed for total PSTs using a receptor binding assay. We then created random forest models to classify shellfish above and below a threshold of toxicity (80 µg 100 g) and used two methods to determine variable importance. We obtained a multivariate model with key variables being sea surface temperature, salinity, freshwater discharge, and air temperature. We then used a similar model trained using lagged environmental variables to hindcast out-of-sample (OOS) shellfish toxicities during April-October in 2017, 2018, and 2019. Hindcast OOS accuracies were low (37-50%); however, we found forecasting using environmental variables may be useful in predicting the timing of early summer blooms. This study reinforces the efficacy of machine learning to determine important drivers of harmful algal blooms, although more complex models incorporating other parameters such as toxicokinetics are likely needed for accurate regional forecasts.