Electrochemical impedance spectroscopy (EIT) is a kind of non-invasive information of the battery, which is intrinsically related to the state, remaining life and health of the battery. In this paper, the real and imaginary data corresponding to all frequencies of EIS were arranged in sequence as frequency features. Based on the neural network model, the frequency features of EIS were taken as input features, and the fitting relationship between battery EIS and battery SOH was constructed. The results show that the root-mean-square error can reach 0.7789. By calculating the importance of each frequency feature in EIS, the Hertz value range of important frequencies is obtained. It is found that high frequency and low frequency of EIS are more important. Through the correlation analysis of important frequency features, it is further obtained that the important frequency features are f61, f65 and f91, and the corresponding frequencies are 20004.453HZ, 7835.48HZ and 17.79613HZ, respectively. The model fitting results show that the fitting effect of the 16 most important frequency features based on EIS is equivalent to or even improved with the fitting effect of all 120 frequency features, and the root-mean-square error is 0.624, which provides a reference for narrowing the detection range of EIS when predicting battery SOH.