Battery　state　of　health　(SOH)　estimation　is　a　critical　but　challenging　demand　in　advanced　battery　management　technologies.　As　an　essential　parameter,　battery　impedance　contains　valuable　electrochemical　information　reflecting　battery　SOH.　This　study　investigates　a　systematic　comparative　study　of　three　categories　of　features　extracted　from　battery　electrochemical　impedance　spectroscopy　(EIS)　in　SOH　estimation.　The　three　representative　features　are　broadband　EIS　feature,　model　parameter　feature,　and　fixed-frequency　impedance　feature.　Based　on　the　deduced　EIS　features,　a　machine　learning　technique　using　Gaussian　process　regression　is　adopted　to　estimate　battery　SOH.　The　battery　aging　and　electrochemical　tests　for　commercial　18650-type　batteries　are　performed,　in　which　the　constant　and　dynamic　discharging　conditions　are　considered　during　battery　aging.　The　battery　life-cycle　capacity　and　EIS　data　are　collected　for　the　machine　learning　model.　The　performance　of　the　constructed　features　is　investigated　and　comprehensively　compared　in　terms　of　estimation　accuracy,　certainty,　and　efficiency.　Experimental　results　highlight　that　using　the　fixed-frequency　impedance　feature　can　realize　outstanding　performance　in　battery　SOH　estimation.　The　average　of　the　maximum　absolute　errors　for　different　cells　under　different　aging　conditions　is　within　2.2%.