Interpretable Diagnostics with SHAP-Rule: Fuzzy Linguistic Explanations from SHAP Values

Abstract

This study introduces SHAP-Rule, a novel explainable artificial intelligence method that integrates Shapley additive explanations with fuzzy logic to automatically generate interpretable linguistic IF-THEN rules for diagnostic tasks. Unlike purely numeric SHAP vectors, which are difficult for decision-makers to interpret, SHAP-Rule translates feature attributions into concise explanations that humans can understand. The method was rigorously evaluated and compared with baseline SHAP and AnchorTabular explanations across three distinct and representative datasets: the CWRU Bearing dataset for industrial predictive maintenance, a dataset for failure analysis in power transformers, and the medical Pima Indians Diabetes dataset. Experimental results demonstrated that SHAP-Rule consistently provided clearer and more easily comprehensible explanations, achieving high expert ratings for simplicity and understanding. Additionally, SHAP-Rule exhibited superior computational efficiency and robust consistency compared to alternative methods, making it particularly suitable for real-time diagnostic applications. Although SHAP-Rule showed minor trade-offs in coverage, it maintained high global fidelity, often approaching 100%. These findings highlight the significant practical advantages of linguistic fuzzy explanations generated by SHAP-Rule, emphasizing its strong potential for enhancing interpretability, efficiency, and reliability in diagnostic decision-support systems.