Comparison of near infrared spectroscopy (NIRS) signal quantitation by multilinear regression and neural networks

Signal quantitation in most near infrared-spectroscopy (NIRS) instruments is achieved through solving simultaneous equations or multiple regression analysis. The aim of this study was to compare NIRS signal quantitation by conventional multiple regression to artificial neural networks. Sixteen adult sheep were used in the study of the effects of changes in cerebral blood flow and metabolism through induction of seizures, ischemia, and hypercapnia. NIRS-derived signal attenuation for relative blood volume (BV) and oxygen desaturation (DESAT) were compared to simultaneous blood flow values measured by laser Doppler flowmetry and venous oxygen-saturation (SVO₂) determined from direct blood gas analysis. The regression for flow provided a zero p-value, a variance S=17.57 and F statistic=50.49. The residuals vs. fits plots suggest that the current model would underestimate values below the mean and overestimate those above the mean. An improved regression model for SvO₂ provided a zero p-value, a variance S=14.1 and F statistic=4.26. Two different neural networks were implemented for flow and oxygen saturation. Both networks "tracked" their values closely and with low cycle errors. Neural networks are powerful tools for evaluation of rapidly changing, variable environments.