2026-06-22T20:19:31.545Z http://ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai

ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai:1:11944 2026-05-27T12:53:11Z 1

AI-driven insights into PCB-138 dynamics in Mediterranean pelagic fish 2025 http://ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai/record/1/11944 https://www.croris.hr/crosbi/publikacija/prilog-skup/909710 T. Bezdan G. Jovanović S. Herceg Romanić B. Mustać M. Perišić Marine environments, particularly such as the semi-enclosed Mediterranean Sea, are long-term sinks for persistent organic pollutants (POPs), including organochlorine pesticides (OCPs) and polychlorinated biphenyls (PCBs). These compounds accumulate in small pelagic fish, essential for both marine ecosystems and human diets. In this study, seven OCPs and 17 PCB congeners were analyzed in sardine, anchovy, round sardinella, chub, and horse mackerel collected over three years in the eastern Mediterranean using high-resolution gas chromatography [1]. Focusing on PCB-138, a toxic, highly chlorinated congener, the study employed a modular artificial intelligence (AI)-based platform [2] identifying key factors influencing PCB-138 levels in fish. The framework integrates advanced machine learning, metaheuristics, and explainable AI leveraging seven ensemble regression algorithms (AdaBoost, CatBoost, ExtraTrees, Gradient Boosting, Histogram Gradient Boosting, LightGBM, XGBoost). Models were evaluated via five-fold cross-validation, and the top three were optimized using Sine Cosine Algorithm and Harris Hawks Optimization to improve predictive accuracy. For model interpretability, SHapley Additive exPlanations (SHAP) and Shapley Additive Global importancE (SAGE) were applied. The ExtraTrees model, optimized with SCA, showed excellent accuracy (mean absolute error=0.0697, mean squared error=0.1604, mean absolute percentage error=0.1394, R2=0.9511). The pollutant abundance order was PCB > DDT > HCH > HCB. Among PCBs, PCB-153 was the most dominant, while p,p’-DDE prevailed among OCPs. Global feature importance (SAGE) reveals that PCB-153 (48.4%), PCB-170 (19.6%), and PCB-180 (11.5%) together accounted for nearly 80%, predictive power, followed by PCB-118 (10.5%), PCB-123 (2.8%), and DDT metabolites (p,p’-DDD: 2.1%, p,p’-DDE: 1.8%). In contrast, HCB, HCH isomers, lower chlorinated PCBs, and biological variables (lipid content, fish size, sampling year) contribute minimally. The results indicate that pollutant load in marine fish is driven mainly by industrial and agricultural pollution, which remains detectable despite long-term bans, rather than biological variability. SHAP analysis reveals that PCB-153, PCB-170, and PCB-180 differ in their influence on PCB-138 concentrations. PCB-153 is the strongest and most variable predictor (SHAP values from –40% to +40%), becoming consistently positive above 1 ng g-1. PCB-170 shows a threshold-dependent impact, negative below 0.1 ng g-1, and strongly positive above while PCB-180 has a weaker, linear effect (SHAP values from –10% to +20%). Interaction analysis shows that PCB-153 and PCB-170 jointly reinforce PCB-138 predictions at higher concentrations, whereas the influence of PCB-153 and PCB-180 interaction remains minimal and stable. These findings suggest stronger co-accumulation between PCB-153 and PCB-170 in marine fish. The findings enhance understanding of POP distribution in marine fish, highlighting the individual and combined effects of specific PCB congeners. They suggest a need to update monitoring strategies to include multiple pollutants and targeted congeners like PCB-170, beyond standard indicator PCBs. Despite data limitations, the scalable AI framework demonstrates strong potential for broader environmental modeling applications. conferenceObject Book of Abstracts The 3rd Conference of the International Association for Biomonitoring of Environmental Pollution – IABEP2025