Machine learning classification of PAH exposure in the Antarctic limpet Nacella concinna

In response to the need for biotic monitoring of petroleum hydrocarbon contamination in Antarctica, we applied machine learning (ML) techniques to classify polycyclic aromatic hydrocarbon contamination in Nacella concinna. Concentrations of 16 priority PAHs were measured in soft tissues of 165 individuals collected on Fildes Peninsula, King George Island. We developed binary classifiers using K-nearest neighbors (KNN), support vector machine (SVM), decision tree (DT), eXtreme gradient boosting (XGB) and random forest (RF), with hyperparameters optimized by grid search and five-fold cross-validation. The findings demonstrate that both RF and XGB exhibited highly stable performance, each achieving an AUC (Area Under Curve) above 0.95. Additionally, applying grid search considerably improved the performance across all models, although the degree of enhancement varied. SHapley Additive exPlanations (SHAP) interpretability analysis further singled out naphthalene as a significant classification marker. The proposed framework demonstrates that combining bivalve bioindicators with tree-based ensemble models enables rapid and reliable screening of petroleum hydrocarbon contamination in Antarctic coastal environments.