New approaches are increasingly being used for studying and evaluating coronary heart disease (CHD), especially since the irruption of metabolomics. The classical approach is to use enzymatically-measured standard lipids and these are still the main markers for assessing risk of CHD. Since metabolomics relies on advanced analytical technologies, such as MS and NMR, using them to estimate standard lipids would be of great interest because there is no need for additional biochemical measures. The present study evaluates partial least squares and N-way partial least squares regression models to predict standard lipid concentrations by using serum and plasma sample sets from various clinical centres. Information provided by editing NMR techniques and 2D diffusion NMR was incorporated in these models using four different data structures. Firstly, the models were calibrated and validated with three of the four sample sets (n = 591) involved. Then the best estimation models were selected and applied to the left-out sample set. This evaluation of a new sample set gave correlation coefficients of predicted versus biochemical variables above 0.86 and %rRMSE lower than 18 %. These values are similar to those found by other studies although, in our case, the results are more general because we used a higher number of samples (n = 785) from different sample sets, different clinical centres and different blood matrices (serum and plasma). Finally, we compared the performance of NMR predicted lipids and enzymatically measured lipids in a clinical case study.