To eliminate the influence of all factors on non-invasive measurement precision of blood components by dynamic spectrum method during data acquisition process, a quality evaluation criterion of dynamic spectrum data needs to be established to improve the stability of the model and precision of the prediction. The number of stable wavelength as a quality evaluation index for the dynamic spectrum data was presented in this article after further analysis of 110 samples, which were all obtained by in-vivo measurements, and 60 samples were picked up as the satisfactory samples. BP artificial neural network was used to establish the calibration model of subjects’ total cholesterol, glucose and hemoglobin values against dynamic spectrum data. The prediction result of the experiment group was improved compared with the control group. The average relative error was decreased from 13.8%, 15.8% and 5.4% to 6.5%, 6.5% and 2.1% respectively, by which the effectiveness of the number of stable wavelength as a quality evaluation index could be proved. By evaluating the quality of the dynamic spectrum data measured, more reliable prediction result can be obtained, which can make the non-invasive measurement of the blood components come to the clinical application sooner.