The accurate estimation of lithium-ion battery state of health (SOH) is essential for system safety and efficiency. This study developed a method that integrates segment-based feature extraction and sparse Gaussian process regression to enhance SOH prediction accuracy. The process entailed segmenting data based on State of Charge (SOC), extracting 32 features per segment, and training distinct GPR models for discharge and charge cycles independently. A comparative evaluation of five Gaussian process regression kernels was performed to assess their predictive power. The proposed method was benchmarked against four alternative machine learning techniques. The test results demonstrated that under high-rate conditions, models based on discharge data between 60% and 70% SOC exhibited superior performance, with an average root mean squared error (RMSE) of 0.26%. In kernel function selection, the Matern 5/2 kernel led to a 31% decrease in maximum RMSE when estimating SOH from charge data, while the rational quadratic kernel resulted in a 64% reduction in RMSE for discharged data.