Assessment of a deep learning model for algal toxin forecasting in coastal Maine

Paralytic shellfish toxins (PST) are detected in seafood globally. Resource managers and seafood producers must make decisions in the face of toxic events in order to protect public health and operate businesses. The coast of Maine in the US sees the occurrence of PSTs annually, and fosters a robust shellfish industry composed of both wild harvest and recently increasing aquaculture. Leveraging monitoring data generated from chemical analysis of PST composition in shellfish, we present a method for grouping together consecutive samples to build a machine-learning training set and make short-term, site-specific predictions towards toxicity of shellfish one week in the future. Over five seasons of operational deployment, based on a process co-developed with end-users, the forecast has consistently achieved >90% accuracy at predicting whether or not sites being currently monitored will surpass the regulatory threshold for PST toxicity.