This paper studies a noninvasive method to measure glucose level based on ultrasonic transducer and near infrared spectrometer. A series pair data of ultrasonic transducer from human finger, palm, wrist and arm are collected six times a day, and 16 spectral data of NIR spectrometer (reflection) from finger are collected by an OGTT experiment. The collected data are calibrated by using partial least squares regression and feed-forward back-propagation artificial neural network to predict the glucose level. In this study, error grid analysis is used to validate the prediction performance. In addition, the accuracy of the calibration models is improved.