To increase the signal-to-noise ratio (SNR) of human near infrared (NIR) spectra, so as to improve the stability and precision of calibration model, the empirical mode decomposition (EMD) method was applied. Eighty-one fingertip absorption curves were collected, with the corresponding clinical examination results obtained immediately. By means of outliers detection and removal, finally 78 samples were determined as the research objects. A three-layer back-propagation artificial neutron network (BP-ANN) model was established and worked for prediction. The results turned out that, through EMD method, the prediction correlation coefficient increased greatly from 0.74 to 0.87. RMSEP was reduced from 12.85 to 8.08 g x L(-1). Other indexes were also obviously improved. The overall results sufficiently demonstrate that it is feasible to use EMD method forhigh SNR pulse wave signals, thus improving the performance of noninvasive hemoglobin calibration models. The application of EMD method can help promote the development of noninvasive hemoglobin monitoring technology.