When measure blood glucose concentration using Mid-IR spectral analysis technique,it is inevitable that high-frequency noises and low-frequency baseline drifting will be added to the spectral data. Thus,it is difficult to extract the weak signal of blood glucose from the obtained spectra. An improved preprocessing method based on wavelet analysis is presented,which can eliminate the noise and correct the baseline drifting at the same time. Firstly,spectra were decomposed into detail and approximation at level J which was estimated according to experience. Then the decomposing level J was determined by further analysis of power density Spectra. The noise was eliminated by filtering high-frequency signal at level J with minimum/maxmum threshold and set high-frequency signal at the other levels to 0. And the baseline at level J was fitted with a quadratic polynomial whose coefficients were calculated by least square curve fitting method. Then remove the quadratic polynomial baseline from the spectral data. After applying this preprocessing method to a set of oral glucose tolerance test data,the cross validation result reveals that the correlation coefficient between prediction value and true value is 0.88,and the root mean square error of prediction is 1.14 mmol/L. The precision of calibration is greatly improved.