Explainable machine learning for predicting shellfish toxicity in the Adriatic Sea using long-term monitoring data of HABs

This study utilizes a 28-year dataset and explainable machine learning techniques, specifically a Random Forest model, to identify key toxic phytoplankton species and environmental factors driving diarrhetic shellfish poisoning in the Gulf of Trieste, thereby enhancing early warning systems for sustainable aquaculture.

The Gist

Imagine the Adriatic Sea as a giant, bustling kitchen where farmers grow mussels for people to eat. But sometimes, invisible "bad actors" (tiny toxic algae) sneak into the water and poison the mussels. If people eat these poisoned mussels, they get sick.

For decades, scientists have been watching this kitchen, taking notes on the water temperature, the river flow, and counting the bad algae. But looking at 28 years of messy, complicated notes is like trying to find a needle in a haystack while wearing foggy glasses.

This paper is about a team of scientists who decided to teach a computer to be the head chef's assistant. They wanted the computer to look at the data and say, "Hey, the mussels are about to get toxic! Close the kitchen!" before anyone gets sick.

Here is the story of how they did it, explained simply:

1. The Problem: The "Black Box" Dilemma

Scientists have long used complex computer programs (Machine Learning) to predict these toxic events. But these programs are often like "Black Boxes." You put data in, and a prediction pops out, but no one knows why the computer made that decision.

If a computer tells a fisherman, "Don't harvest today," the fisherman needs to know why. Is it because the water is too salty? Because a specific algae is present? Without an explanation, the fisherman might not trust the computer, and they might lose money or, worse, people might get sick.

2. The Solution: The "Explainable" Detective

The team built a new computer model that doesn't just guess; it explains its reasoning. Think of it like a detective who not only solves the crime but also shows you the clues on the whiteboard.

They fed the computer 28 years of data, including:

• The Bad Guys: Counts of specific toxic algae (like Dinophysis fortii).
• The Weather: Rain, wind, and temperature.
• The Water: How salty it is and how much fresh water is flowing in from the Soča River.

3. The Training: Teaching the Computer

The data was tricky. Most of the time, the mussels were safe (the "negative" class). Only about 12% of the time were they toxic (the "positive" class). It's like trying to teach a dog to bark only when a specific, rare bird flies by, but the bird only shows up once a month.

To fix this, the scientists used a clever trick called SMOTE. Imagine you have a pile of 100 safe mussels and only 12 toxic ones. To teach the computer better, they created "synthetic" examples of the toxic ones (like making photocopies of the rare clues) so the computer could learn what danger looks like without getting confused by the overwhelming number of safe days.

4. The Champion: The "Random Forest"

They tried four different types of computer brains:

• Decision Tree: A simple flowchart (If A, then B). Easy to understand, but not very smart.
• Neural Network: A complex brain that mimics the human mind. Very smart, but a "Black Box."
• Support Vector Machine: A mathematical divider.
• Random Forest: A committee of many simple decision trees that vote on the answer.

The winner was the Random Forest. It was the best at predicting toxicity. But here's the magic: the team didn't just stop there. They used special tools (called SHAP and Permutation Importance) to ask the Random Forest, "Why did you vote 'Toxic'?"

5. The "Aha!" Moments: What the Computer Learned

The computer's explanation was surprisingly clear and matched what human experts already knew, which made everyone trust it more.

• The Main Culprit: The computer identified Dinophysis fortii as the biggest warning sign. If this specific algae is present in high numbers, the mussels are likely toxic.
• The Weather Connection: It learned that low salinity (water that isn't very salty) and freshwater flow from the river are key triggers.
  • Analogy: Think of the river like a giant hose. When it rains and the river flows fast, it dumps fresh water into the sea. This fresh water sits on top of the salty sea water, creating a "stratified" layer (like oil on water). This calm, layered environment is a perfect party for the toxic algae to grow.
• The Season: The computer confirmed that late summer and autumn (September to November) are the most dangerous times.

6. Why This Matters

This isn't just about math; it's about trust and safety.

• For the Fishermen: Instead of waiting for a lab test (which takes days) to confirm toxicity, they can get an early warning. "The computer sees the algae and the river flow; it's time to stop harvesting." This saves them money and keeps their reputation safe.
• For the Public: It means safer seafood on the table.
• For Science: It proves that we don't have to choose between a "smart" computer and a "understandable" one. We can have both.

The Bottom Line

The scientists built a digital crystal ball that doesn't just predict the future; it tells you why the future is happening. By understanding that specific algae + river floods + warm weather = danger, they created a tool that helps protect the Adriatic Sea's seafood industry from invisible threats.

It's like having a weather forecaster who doesn't just say "It will rain," but explains, "It will rain because the clouds are heavy, the wind is blowing from the west, and the ground is already wet." That's the power of Explainable AI.

Technical Summary

Here is a detailed technical summary of the paper "Explainable machine learning for predicting shellfish toxicity in the Adriatic Sea using long-term monitoring data of HABs."

1. Problem Statement

The study addresses the critical challenge of predicting Diarrhetic Shellfish Poisoning (DSP) events in mussels (Mytilus galloprovincialis) within the Gulf of Trieste (Adriatic Sea).

• Context: Harmful Algal Blooms (HABs), particularly those caused by dinoflagellates like Dinophysis, pose significant risks to human health and the aquaculture industry.
• Current Limitations: Existing monitoring systems rely on reactive measures (closing harvests only after toxicity is confirmed), leading to economic losses and periods of uncertainty. While Machine Learning (ML) has been used to predict HAB species occurrence, predicting the actual toxicity of shellfish is more complex and less explored.
• The "Black Box" Issue: Many high-performing ML models lack interpretability. For regulatory bodies and stakeholders to trust and adopt ML-based Early Warning Systems (EWS), it is crucial to understand the causal drivers and decision logic behind the predictions.

2. Methodology

Data Acquisition and Preprocessing

• Dataset: A newly created, open-access dataset spanning 28 years (1994–2021).
• Variables:
  • Biological: Abundance of five key DSP-producing species (D. fortii, D. caudata, D. sacculus, D. tripos, P. rotundatum) and total DSP species (DSP-tot).
  • Toxicity: Binary target variable (Positive/Negative) based on regulatory limits (176 µg kg⁻¹). Data sources included Mouse Bioassay (1994–2013) and LC-MS/MS (2014–2021), standardized to binary outcomes.
  • Environmental: Sea Surface Temperature (SST), salinity, river discharge (Soča River), precipitation, air temperature, wind speed, and solar irradiance.
• Preprocessing Pipeline:
  • Alignment: Mismatched sampling frequencies were resolved by assigning toxicity results to phytoplankton observations within a 30-day window.
  • Imputation: Missing environmental data (38% of values) were interpolated using nearest-neighbor measurements within a 30-day window.
  • Dimensionality Reduction: UMAP was used to visualize data structure, revealing significant overlap between positive and negative classes.
  • Class Imbalance Handling: The dataset was heavily imbalanced (88% negative, 12% positive). The authors employed a pipeline combining SMOTE (Synthetic Minority Over-sampling Technique) and Edited Nearest Neighbors (to remove conflicting negative instances) to balance the training set.
  • Final Dataset: 877 instances (745 negative, 132 positive) split 70/30 for training/testing.

Model Training and Evaluation

• Algorithms Tested: Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN/MLP).
• Optimization: Hyperparameters were tuned using Grid Search with stratified 5-fold cross-validation. The optimization metric was the F1 score to balance precision and recall.
• Evaluation Protocol: To ensure statistical reliability, the entire pipeline was run 100 times, averaging performance metrics (Precision, Recall, F1-score).

Explainable AI (XAI) Integration

To address the "black box" nature of complex models:

• Decision Trees: Visualized directly to inspect decision rules.
• Random Forest (RF): Analyzed using:
  • Permutation Importance: To rank feature relevance by measuring performance drop upon shuffling.
  • SHAP (SHapley Additive exPlanations): Used both globally (Beeswarm plots) to show feature impact distribution and locally (Force plots) to explain individual predictions.

3. Key Results

Model Performance

• Best Performer: Random Forest (RF) achieved the highest and most stable performance.
  • Average Precision: 0.74 (±0.09)
  • Average Recall: 0.59 (±0.08)
  • Average F1 Score: 0.65 (±0.07)
• Comparison: DT and ANN performed similarly (F1 ≈ 0.43), while SVM was less effective.
• Trade-off: The models were optimized for F1 score, but the study notes that for real-world EWS, prioritizing Recall (detecting all positive events) is often more critical, even if it lowers Precision (increasing false positives).

Explainability Insights

• Key Predictors: Both Permutation Importance and SHAP identified the same top drivers:
  1. Dinophysis fortii abundance: The single most influential variable.
  2. Dinophysis caudata abundance.
  3. Total DSP species abundance (DSP-tot).
  4. Environmental Factors: Soča River flow, Salinity, and Precipitation.
• Mechanism:
  • High abundance of D. fortii and D. caudata strongly increases the probability of a positive toxicity prediction.
  • Salinity and River Flow: Lower salinity and higher river discharge (freshwater input) correlate with higher toxicity risk, likely due to water column stratification favoring Dinophysis growth.
  • Decision Rules (DT): The model identified a specific threshold: D. fortii > 30 cells L⁻¹ predicts toxicity. If below this, the presence of D. caudata combined with salinity ≤ 36.17 predicts toxicity.

4. Key Contributions

1. Dataset Creation: Generated and openly released a comprehensive 28-year dataset linking toxic phytoplankton, environmental parameters, and mussel toxicity in the Adriatic Sea.
2. Methodological Advancement: Successfully applied ML to the difficult task of predicting shellfish toxicity (rather than just species presence) in a marine environment.
3. Explainability Framework: Demonstrated how XAI techniques (SHAP, Permutation Importance) can bridge the gap between complex ML models and regulatory decision-making, identifying specific biological and physical drivers of DSP outbreaks.
4. Practical EWS Potential: Provided a validated, interpretable model that can serve as a cost-effective component of Early Warning Systems, allowing for proactive rather than reactive management of shellfish harvests.

5. Significance and Future Outlook

• Significance: This study moves beyond "black box" predictions in marine ecology. By proving that ML models can be both accurate and interpretable, it builds the necessary trust for authorities to integrate these tools into official monitoring protocols. The findings confirm that D. fortii is the primary driver of DSP in the region, validating biological hypotheses with data-driven evidence.
• Limitations: The model's generalizability is currently limited to the specific conditions of the Gulf of Trieste due to the small size of the positive class and the specific hydrodynamic characteristics of the region.
• Future Work: The authors suggest expanding the dataset with more positive toxicity events, improving temporal resolution, and incorporating data from neighboring regions to enhance model robustness. They also recommend calibrating models specifically for high-recall scenarios in operational EWS.