Abstract: To ensure the safety, reliability and longevity of battery systems, accurate estimation of the State of Health (SOH) of lithium-ion batteries is essential. As an internal state variable, SOH is difficult to measure directly with sensors and is therefore often estimated through indirect methods. The accuracy of SOH estimation largely depends on the quality of the extracted health features that are correlated with battery aging. This review systematically analyzes and evaluates mainstream feature extraction methodologies for lithium-ion battery SOH estimation. It clarifies the link between macroscopic aging phenomena (capacity fade and impedance rise) and microscopic electrochemical degradation mechanisms, such as loss of active material (LAM) and loss of lithium inventory (LLI). A comprehensive survey is conducted on five primary feature categories: (1) Voltage-current curve features, derived from standard charging protocols (e.g., Constant Current-Constant Voltage, CC-CV), including temporal indicators and capacity metrics within specific voltage windows. (2) Differential curve features, such as Incremental Capacity Analysis (ICA) and Differential Voltage Analysis (DVA), identifying electrochemical phase transitions whose peak attributes (height, position, area) serve as health indicators. (3) Pulse power characterization features, obtained from Hybrid Pulse Power Characterization (HPPC) tests, reflecting DC internal resistance (DCR) and variations in the open-circuit voltage (OCV) versus state of charge (SOC) curve. (4) Electrochemical impedance spectroscopy (EIS) features, extracted from raw impedance data, including parameters fitted using equivalent circuit models (ECM) and deconvolution results from distribution of relaxation times (DRT) analysis. (5) Multi-physics field features, which utilize non-electrical signals from thermal, ultrasonic, and mechanical sensors, providing additional diagnostic dimensions. Publicly available datasets (e.g., NASA, CALCE, Oxford) are also reviewed as benchmarks. The analysis finds that voltage-current curve features are computationally efficient but typically require full charging cycles. While ICA/DVA offer deep mechanistic insight by linking peak changes to LAM and LLI, their susceptibility to noise and current rate complicates online implementation. HPPC-derived features effectively track impedance growth but require accurate OCV correction. EIS provides the most comprehensive diagnostic information, with ECM offering physically meaningful parameters and DRT excelling at decoupling overlapping processes, though measurements are time-intensive. Multi-physics features capture structural and thermal degradation, offering complementary perspectives. A key finding is that no single feature can reliably provide robust and high-precision SOH estimation under complex and variable real-world conditions. Given the limitations of single features, future research is expected to focus on: (1) establishing standardized public benchmarks and evaluation protocols to enable objective comparison and accelerate technological progress; (2) fusing multi-physics features (electrical, thermal, mechanical) to develop more comprehensive and robust health indicators; and (3) integrating physical models with data-driven methods, such as physics-informed neural networks (PINNs), to enhance model interpretability, data efficiency, and generalization.