Comparison of analytical mathematical approaches for identifying key nuclear magnetic resonance spectroscopy biomarkers in the diagnosis and assessment of clinical change of diseases.

Nuclear magnetic resonance (NMR) spectroscopy is a rapidly emerging technology that can be used to assess tissue metabolic profile in the living animal. At the present time, no approach has been developed 1) to systematically identify profiles of key chemical alterations that can be used as biomarkers to diagnose diseases and to monitor disease progression; and 2) to assess mathematically the diagnostic power of potential biomarkers. To address this issue, we have evaluated mathematical approaches that employ receiver operating characteristic (ROC) curve analysis, linear discriminant analysis, and logistic regression analysis to systematically identify key biomarkers from NMR spectra that have excellent diagnostic power and can be used accurately for disease diagnosis and monitoring. To validate our mathematical approaches, we studied the striatal concentrations of 17 metabolites of 13 R6/2 transgenic mice with Huntington's disease, as well as those of 17 wild-type (WT) mice, which were obtained via in vivo proton NMR spectroscopy (9.4 Tesla). We developed diagnostic biomarker models and clinical change assessment models based on our three aforementioned mathematical approaches, and we tested all of them, first, with the 30 original mice and, then, with 31 unknown mice. Their prediction results were compared with genotyping-the gold standard. All models correctly diagnosed all of the 30 original mice (17 WT and 13 R6/2) and all of the 31 unknown mice (20 WT and 11 R6/2), with a positive likelihood ratio approximating infinity [1/0 (→ ∞)], and with a negative likelihood ratio equal to zero [0/1 = 0].