Validation of machine learning models for heavy metals bioavailability prediction: A comparative study

Heavy metals (HMs) bioavailability in composting poses significant environmental and health risks, necessitating accurate predictive models for effective risk assessment and management. Traditional models and existing machine learning (ML) models have shown promise, but still suffer from limitations such as poor generalizability, dataset inconsistencies, and lack of uncertainty quantification. To address these limitations, this study validated and compared the performance of four advanced ML models that encompassed Gaussian Process Regression (GPR), Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Extreme Learning Machine (ELM) in predicting the bioavailability of HMs (Arsenic (As), Cadmium (Cd), Copper (Cu), Chromium (Cr), and Zinc (Zn)). The models were evaluated using several statistical metrics. Results revealed that the GPR model consistently outperformed all other models, achieving consistently higher overall determination coefficient (R2) values of 0.972, 0.938, 0.841, 0.925, and 0.956 for As, Cd, Cr, Cu, and Zn respectively. Supported by the lowest RMSE and MAE across all metals. Alternatively, the RF Model excelled in Cd, while XGBoost and ELM models exhibited greater variability and lower predictive stability. These findings signified the superiority of the GPR model in HMs bioavailability prediction, making it the preferred choice for environmental monitoring and policymaking. However, challenges remain in the generalizability of ML models across different environmental settings. Future research could validate these models using independent datasets from different geographic regions and explore hybrid modeling methods to further improve predictive accuracy and adaptability in real-world applications.