Treatment of diabetes mellitus is a public health related problem of modern healthcare. Surveys show that current methods to estimate the required amount of insulin are quite inefficient in practice as they are based on experience. This paper offers a new approach to predict the glucose level of people with diabetes. It combines two efficient models of the literature: one for nutrient absorption and one for glucose control. The combination of them tracks the blood sugar level considering nutrition composition, applied insulin and initial glucose level. Compared to already existing mixed meal models, the current version takes into account a more detailed nutrition composition (protein, lipid, monosaccharide, fiber and starch) supported by our expert dietary systems. Although the model gives satisfactory results even with parameter sets taken from literature, parameter training by genetic algorithms yields a better tracking of the patients.