Interpreting is a cognitively demanding task requiring the simultaneous management of multiple processes under severe time constraints. While quantifying interpreting difficulty is essential for training and assessment, existing methods struggle to reconcile predictive accuracy with explanatory power. To bridge this gap, this study introduces a novel, interpretable evaluation framework based on Kolmogorov-Arnold Networks (KAN). Unlike black-box deep learning models, KAN employ learnable univariate functions on network edges, enabling the direct extraction of explicit symbolic formulas from data. Our model achieved superior predictive accuracy (R² = 0.9893), reducing the Mean Squared Error (MSE) by 49.8% compared to a standard BP Neural Network. Global sensitivity analysis of the derived formula revealed key quantitative laws governing difficulty, most notably a non-linear threshold effect for speech rate and a synergistic composite effect between noise and context dependency. These findings provide a transparent, data-driven foundation for cognitive load theories in interpreting and offer practical implications for developing intelligent, adaptive interpreter training systems.