Empirical mode decomposition (EMD) algorithm combined with the theory of dynamic spectrum extraction at frequency domain was applied to the noninvasive measurement of hemoglobin concentration. Fifty seven cases’ photoplethysmography was collected in the range of 636.98～1 086.86 nm in vivo. After the denoising preprocess through the EMD method for each wavelength pulse wave of each sample separately, dynamic spectrum of each sample was made up of all peaks extracted by Fourier transform. Partial least squares regression model was used to establish the calibration and prediction of hemoglobin concentration. Compared to the modeling results without EMD, the correlation coefficient of predicted values and the real values was increased from 0.879 8 up to 0.917 6. The root mean square error of prediction set was reduced from 6.675 9 to 5.300 1 g·L-1 and the relative error was reduced from 8.45% to 6.71%. The modeling accuracy has been greatly improved. The results showed that EMD algorithm can be effectively applied to denoise the spectral data and improve the accuracy of the non-invasive measurement of blood components.