An accurate prediction model plays a very important role in the quantitative spectrum analysis. Aiming at the problem of large model prediction error caused by information lacking and imbalanced information distribution in small sample set space, in this paper, based on traditional MDMTD (multidistribution mega trend diffusion) method, a HybridMega Trend Diffusion (HybridMTD) technique is proposed to construct virtual sample space, which further expends the training sample set and improves the information distribution of the sample set space, and then obviously reduces model prediction error. The spectrum data sets of total cholesterol and triglyceride in whole blood samples were utilized to carry out comparison and experiment verification. The experiment results show that the PLS prediction models established based on the reconstructed data set with virtual samples added can provide lower mean prediction mean square error MRmesp (mean of RMSEP). The values of MRmesp of total cholesterol and triglyceride are 041 and 045 mmol/L, respectively. Compared with traditional MDMTD method, the errors are reduced by 467% and 224%, respectively. The proposed HybridMTD method can construct an adequate number of highquality virtual samples; the prediction model corresponding to the sample set with the virtual samples filled obviously reduces the prediction error, and has superior prediction performance compared with the existing MTD method. The application of HybridMTD technique in blood spectrum analysis effectively enhances the evaluation quality of physiological indicators, speeds up screening speed for cardiovascular disease and reduces its risk.