Accurate and interpretable estimation of the State of Health (SOH) of lithium-ion batteries is essential for the safety and efficiency of electric vehicles and energy storage systems. While machine learning (ML) models have shown strong performance in SOH prediction, their black-box nature limits practical deployment. This paper presents a comparative study of multiple explainable AI (XAI) methods – SHapley Additive exPlanations (SHAP), Local Interpretable Model-Agnostic Explanations (LIME), Integrated Gradients (IG), and DeepLIFT – applied to four standard ML models SVM, Bayesian Ridge Regression, GPR, and ANN trained on NASA’s battery cycle dataset. The objective is to systematically evaluate and contrast these XAI techniques in terms of consistency, stability, and feature relevance for SOH interpretation. Our comparative analysis reveals that SHAP demonstrates superior attribution consistency with LIME (ρ=0.89) and highest stability under perturbation (variance =0.00021), while gradient-based methods show moderate agreement (ρ=0.68-0.74) but greater sensitivity. LIME achieves high surrogate fidelity (R2 > 0. 9 3) for linear models but exhibits lowest robustness (variance =0.00112). Our systematic evaluation framework quantifies critical trade-offs between explanation reliability and computational efficiency, demonstrating that perturbation-based methods outperform gradient-based approaches in stability metrics by 80 %. This work provides the first comprehensive quantitative benchmark for XAI method selection in battery diagnostics and contributes actionable guidance for deploying interpretable SOH estimation in safety-critical applications.