With the change in their social environment and life patterns, the eating habits of modern people have become more diverse. Eating habits are closely related to health, and diet management for individual healthcare is required in this era characterized by poor health and increased longevity. In this paper, we propose a knowledge-based hybrid decision model for nutrition management that uses neural networks. The proposed method is a food recommendation model to help users make dietary nutrition-related decisions on a health platform. It is a hybrid recommendation method that considers both physical and mental health. It selects foods that are positively related to users’ physical health as candidates and predicts users’ preferences through adaptive learning. A previously developed dietary nutrition service ontology is used to select foods that appear to affect the user’s health positively. Conventional preference prediction methods include collaborative, content-based, knowledge-based, and image-based filtering. These methods use a hybrid model or machine learning, data mining, and artificial intelligence methods to compensate for the disadvantages of each filtering type. For preference prediction, healthcare and food preference data are collected in an on/off line environment. The data consist of age, sex, body mass index, region, chronic disease, and food preferences. Food preferences include the dietary nutritional components of food, which makes it possible to infer the user’s preferences for foods containing calories, carbohydrates, protein, fat, sugars, sodium, cholesterol, saturated fatty acids, and trans fatty acids. The user’s preference for food is composed of output variables, and other variables are composed of input variables. The variables consist of 11 healthcare data variables, 2 preference data variables, 10 dietary nutrition data variables, 22 input variables, and 1 output variable. The variables that we constructed are used to arrange transactions and supervised learning is conducted in a neural network structure. In total, 3152 transactions, 80% of the collected data, were used as learning data and 788 transactions, 20% of the collected data, as test data. Using the test data, we evaluated the performance of four prediction models based on a learned neural network, user correlation, average replacement, and regression analysis, respectively. The result of the performance evaluation showed that the proposed method was superior to the conventional method in that it solved the cold-start and the sparsity problem. In addition, the user’s satisfaction evaluation result was 3.92 on a five-point scale, showing overall satisfaction. Therefore, on the platform it is possible to recommend dietary nutrition for people suffering chronic diseases according to their lifestyle and in consideration of their health status and preferences. The platform selects a suitable candidate food according to the health condition of the user and provides a recommendation for N foods using the Top-N of the user’s food preferences.