Prediction and Risk Warning of BMWP Index of Tropical Rivers Based on Machine Learning

Global river ecosystems face growing threats from pollution and habitat degradation, creating an urgent need for efficient and interpretable health assessment tools. The Biological Monitoring Working Party (BMWP) index is widely used for water quality evaluation but suffers from low efficiency and subjectivity due to reliance on expert judgment and limited ecological representation. To improve its application, we analyzed 161 benthic macroinvertebrate samples from the Key River Basins in Hainan Province, China, and developed several machine learning models, including LASSO regression, to predict BMWP values. The LASSO model outperformed others, achieving R2 = 0.88, demonstrating strong stability and generalization. SHAP analysis identified the Shannon diversity index, number of sensitive taxa, and Hilsenhoff index as key drivers of BMWP variation: higher diversity and sensitive taxa indicated better water quality, while the Hilsenhoff index reflected higher pollution risks. Based on these insights, BMWP values were classified into five risk levels, each corresponding to distinct ecological conditions, with significant differences in indicator distribution supporting their ecological basis. Spatial analysis revealed high-risk clusters in the upper and lower reaches of the Changhua River, a downstream gradient along the Nandu River, and overall better conditions in the Wanquan River. These findings highlight priority areas for ecological management. The integrated framework—combining prediction, interpretability, and early warning—provides a systematic approach for ecological risk identification and spatial conservation planning in tropical river systems.