Limitations of SHAP-based interpretations in environmental and membrane filtration applications

Maliwan et al. (2025) identified key parameters in microplastic ultrafiltration using interpretable machine learning (SHAP), attributing 57.6-70.6 % feature importance to factors like transmembrane pressure. This paper critically examines their methodological approach, highlighting significant concerns regarding SHAP's application. SHAP values are inherently model-dependent and lack ground truth for validating feature importance accuracy, leading to potentially biased and erroneous conclusions; high prediction accuracy does not ensure reliable insights. SHAP's underlying assumptions, particularly feature independence, rarely hold in complex environmental systems characterized by multicollinearity, potentially misattributing variable importance. We advocate for a more robust analytical framework incorporating unsupervised machine learning (e.g., feature agglomeration) and nonlinear nonparametric statistical methods (e.g., Spearman's correlation) to provide more reliable insights into variable relationships, moving beyond model-dependent interpretations.