Comparative chemometric analysis of transverse low-field 1H NMR relaxation data

Transverse relaxation data from low-field (23.2 MHz) H-1 NMR was investigated by different data analytical approaches. The quantitative information in the relaxation data of 200 intact salmon samples with respect to overall water and fat content was evaluated by four different data analytical methods: (a) bi-exponential curve fitting followed by linear regression, (b) forward selection of variables followed by multiple linear regression (FS-MLR), (c) partial least squares regression (PLS) and (d) non-negative alternating least squares regression (NN-ALSR). The investigation demonstrates that the quantitative prediction performance is significantly enhanced (reduction of the prediction error from 14% to 34%) by the use of multivariate chemometric procedures such as PLS. While PLS is an extraordinarily robust and efficient algorithm its strictly orthogonal latent variables suffer from a difficult qualitative interpretation. NN-ALSR and FS-MLR which exhibit a quantitative performance comparable to PLS do not suffer from this problem, but are more unstable and ineffective data analytical techniques.