If temporal variation of a person's hunger degree could be estimated, it would be possible to adjust his/her eating habits and/or prevent obesity. It is well-known that there is a negative correlation between a hunger degree and a blood glucose level. However, it is hard to measure a person's blood glucose level anytime and anywhere, because it relies usually on an invasive method (e.g., blood sampling). This paper proposes a method for estimating a person's hunger degree in a non-invasive way. Our proposed method is composed of (1) a blood glucose level estimation model based on logs of meals and exercises, and (2) a hunger degree estimation model based on the estimated glucose level. The former model is constructed by correlating an actual blood glucose level and logs of meals and exercises with a machine learning technique. Here, the actual blood glucose level is measured by a commercial blood glucose meter invasively. The latter model is constructed by associating the measured blood glucose level with a subjective hunger degree. We also design and develop a mobile application for facilitating a user to easily record meals and exercises information. Through an experiment with a subject, we confirmed that our system can estimate a blood glucose level within about 14% mean percentage error and finally estimate hunger degree within about 1.3 levels mean error among 10 levels.