Decoding fish health trajectories across anthropogenically impacted habitats: integrating machine learning approaches for ecological assessment in Pakistan

Abstract

This study examines the flow and transformation of energy and nutrients using a machine learning-based approach to design scientific experiments, improve precision in aquaculture predictions, and reduce reliance on experimental animals. The model predicts measurable aquaculture-related parameters, including morphometric patterns, macronutrient composition, stress levels, and toxicity. Thus, the Random Forest (RF) ensemble learning approach in modeling complex biological responses in aquatic ecotoxicology reflects the physiological condition of fish across different environments and study sites. The highest average dimensions were observed in specimens from Waste Water sites (III and IV), with a condition factor of 1.49. Proximate composition analysis reveals a significant decline in nutritional quality at these sites, with protein levels dropping to a minimum of 14.56% and lipids to 2.62%, associated with heavy metal bioaccumulation and related oxidative stress, which often results in increased water content (79.87%) at the expense of muscle tissue, following an inverse relationship. While homeostatic control restricted the predictive utility of essential elements, the RF model detected differences among metals and oxidative stress drivers (SOD and CAT) in the first decision nodes at 95.1% cross-validation accuracy. Furthermore, the RF framework automatically categorized geographical risks without regulatory cut-offs by accounting for energy reallocation characteristics (protein or so). Overall, our findings support the Random Forest approach as a high-precision machine learning technique for aquatic watershed risk-indexing and multi-tier biomonitoring. The proposed model could be valuable for fish researchers, scientists, technicians, and farm operators seeking to improve learning and adaptation to new technologies.