The rapid quantification method of human serum glucose was established by using the Fourier transform infrared spectroscopy（FTIR） and attenuated total reflection（ATR）.The subtracted spectrum between glucose solution and water was calculated which was 1 200-900 cm-1 and named fingerprint region.Both partial least squares（PLS） models of whole spectrum（4 000-600 cm-1） and fingerprint spectrum were built.The predicted result of fingerprint region was better than that of the whole spectral.Three PLS models for all samples，low concentration samples and high concentration samples were built.Based on this,compound decision of three models was carried on.Finally，the independent test set was performed for model verification.The results showed that the prediction bias of the combination model for human serum glucose was obviously better than that of the whole spectral or the fingerprint region，and for the samples whose value were nearby threshold，the difference of prediction errors for low concentration model and high concentration model was small.RMSEP and RP were 0.732 mmol/L and 0.948，respectively，by the concentration section model.