Identifying environmental factors associated with tetrodotoxin contamination in bivalve mollusks using eXplainable AI

TL;DR

This study tackles the prediction of tetrodotoxin contamination in bivalve mollusks by leveraging an explainable LSTM trained on Dutch Zeeland data with a 35-day environmental window. SHAP-based explanations reveal that day length-related factors (sunrise, sunset, sun hours) and hydroclimatic variables (global radiation, water temperature, chloride concentration) are key drivers of TTX presence. The approach achieves strong discriminative performance (AUC ~0.91–0.93) and remains robust under sensitivity analyses, offering a practical tool for authorities and industry to target monitoring efforts. The work demonstrates how temporal deep learning combined with XAI can illuminate environmental drivers of marine toxins and support proactive risk management.

Abstract

Since 2012, tetrodotoxin (TTX) has been found in seafoods such as bivalve mollusks in temperate European waters. TTX contamination leads to food safety risks and economic losses, making early prediction of TTX contamination vital to the food industry and competent authorities. Recent studies have pointed to shallow habitats and water temperature as main drivers to TTX contamination in bivalve mollusks. However, the temporal relationships between abiotic factors, biotic factors, and TTX contamination remain unexplored. We have developed an explainable, deep learning-based model to predict TTX contamination in the Dutch Zeeland estuary. Inputs for the model were meteorological and hydrological features; output was the presence or absence of TTX contamination. Results showed that the time of sunrise, time of sunset, global radiation, water temperature, and chloride concentration contributed most to TTX contamination. Thus, the effective number of sun hours, represented by day length and global radiation, was an important driver for tetrodotoxin contamination in bivalve mollusks. To conclude, our explainable deep learning model identified the aforementioned environmental factors (number of sun hours, global radiation, water temperature, and water chloride concentration) to be associated with tetrodotoxin contamination in bivalve mollusks; making our approach a valuable tool to mitigate marine toxin risks for food industry and competent authorities.

Figures (6)

• Figure 1: Receiver operating characteristics of above Action Limit (AL) classification by our long short-term memory (LSTM) method on the validation set (N=154, area under the curve (AUC)=0.91, solid line), and test set (N=144, AUC=0.93, dotted line). The 95% Confidence Interval (CI) (AUC=0.79-0.98, shaded area) was obtained by bootstrapping the validation set 10,000 times. The performance on the test set shows similar results to the validation set. There appears to be one TTX positive sample in the test set, which the model fails to correctly classify. Finally, at a threshold of 0.57, selected based on a validation sensitivity >90%, the test set obtained a sensitivity (Sens) of 93% and a specificity (Spec) of 81% (diamond icon).
• Figure 2: Explainable AI (XAI) shows, with SHaply Additive exPlanations (SHAP), the global explanations for the long short-term memory model. The most important features are time of sunrise, time of sunset, sun hours, chloride concentration, global radiation, and water temperature. Each of these features were shown to be significant ($p < 0.05$) for the model’s prediction, with the exception being water temperature ($p = 0.10$).
• Figure 3: Difference in average SHAP values between samples with a TTX concentration below the action limit (green bars, left) and above the action limit (blue bars, right) for the six most important features according to our XAI method. Each of these features was shown to be significant ($p < 0.05$) for the model’s prediction, except water temperature ($p = 0.10$).
• Figure 4: Left: The SHAP values of the false negative identified in \ref{['ROCAL']}. Right: a correctly classified true positive sample is shown. Besides each SHAP value, the average SHAP values of all TTX positive samples are shown. The false negative sample diverges in prediction direction on all features, except water temperature.
• Figure 5: Receiver operating characteristics of above Legal Limit (LL) classification by our long short-term memory (LSTM) method on the validation set (N=154, area under the curve (AUC)=0.93, solid line), and the test set (N=144, AUC=0.94, dotted line). The 95% Confidence Interval (CI) (AUC=0.84-0.98, shaded area) was obtained by bootstrapping the validation set 10,000 times. As with the main analysis (\ref{['ROCAL']}), the sensitivity analysis showed a test set performance similar to the validation set. Finally, at a threshold of 0.57, selected based on a validation sensitivity $>$90%, the test set obtained a sensitivity (Sens) of 100% and specificity (spec) of 78% (diamond icon).