Automatic data reduction and pattern recognition methods for analysis of 1H nuclear magnetic resonance spectra of...

Multivariate data analysis techniques have been used to compare 600-MHz 1H nuclear magnetic resonance (NMR) spectra of urine obtained from patients with inborn errors of metabolism (IEM) and urine obtained from healthy subjects. These spectra are very complex; each contains many thousands of resonances with a high dynamic range. A consistent method of reducing this wealth of data to manageable proportions is presented as a two-stage process. Computer-based spectral descriptors are automatically generated and then reduced to two-dimensional maps for visualization of clustering. Data-scaling methodology has been developed to achieve complete separation between spectra from control adults and those from adult patients with independently diagnosed IEM. The methods were refined by relating IEM samples to the mean of the control samples and applying supervised learning techniques to identify descriptors contributing to class separation. This approach allowed separation of the various classes of IEM and achieved optimal separation of patients with cystinuria from those with oxalic aciduria; the principal metabolites responsible for this separation were determined as lysine and glyoxalate. The methods developed were then extended by application to the more subtle problem of classifying urine collected from healthy subjects under different physiological conditions (i.e., pre- and post-exercise and in different stages of hydration) where, unlike the IEM case, any underlying biochemical differences were not known at the outset. Fluid-loaded and fluid-deprived samples could be partially separated as well as fluid-deprived and fluid-restored samples. Partial classification of samples on the basis of subject was also observed. Therefore, intersubject differences were liable to obscure the separation by physiological state. However, by relating each sample to a mean of the normal daily urine samples for the same person and applying a form of "range scaling" to exclude data which contributed least to class separation, improved classification of the hydration states resulted, from which it was possible to deduce those biochemical substances which were altered. These novel techniques for the data reduction and classification of NMR spectra make comprehensive use of all of the NMR spectral information and have clear potential to assist in clinical diagnosis.