NMRES: an artificial intelligence expert system for quantification of cardiac metabolites from 31phosphorus nuclear magnetic resonance spectroscopy

Abstract

The application of high-resolution 31Phosphorus Nuclear Magnetic Resonance (31P NMR) Spectroscopy in biology and medicine has provided new insights into biochemical processes and also a unique assessment of metabolites. However, accurate quantification of biological NMR spectra is frequently complicated by: (a) non-Lorentzian form of peak lineshapes, (b) contamination of peak signals by neighboring peaks, (c) presence of broad resonances, (d) low signal-to-noise ratios, and (e) poorly defined sloping baselines. Our objectives were to develop an expert system that captures and formalizes 31P NMR spectroscopists' expert knowledge, and to provide a reliable, efficient, and automated system for the interpretation of biological spectra. The NMR Expert System (NMRES) was written in the C and OPS5 programming languages and implemented on a Unix-based (Ultrix) mainframe system with XWindows bit-map graphics display. Expert knowledge was acquired from NMR spectroscopists and represented as production rules in the knowledge base. A heuristic weights method was employed to determine the confidence levels of potential peaks. Statistical and numerical methods were used to facilitate processing decisions. NMR spectra obtained from studies of ischemic neonatal and immature hearts were used to assess the performance of the expert system. The expert system performed signal extraction, noise treatment, resonance assignment, intracellular pH determination, and metabolite intensity quantitation in about 10 s per 4 KB (kilobyte) spectrum. The peak identification success rate was 98.2%. Peak areas and pH estimated by the expert system compared favorably with those determined by human experts. We conclude that the expert system has provided a framework for reliable and efficient quantification of complex biological 31P NMR spectra.