International Classification of Diseases (ICD) coding is an internationally unified diagnostic system, which sets a unique code for each patient who carries particular disease. The primary diagnosis indicates the most crucial disease for the patient in the hospitalization, and its coding is very important for both the patient and the hospital. The current ICD encoding methods mostly use text as the input of the model, and the unstructured nature of text data results in a relatively scattered feature distribution, which is not conducive to feature extraction and interpretability research of the models. In this paper, we utilized a knowledge graph to transform text into graph structured data, and used an improved graph convolutional model to extract features from the transformed graph, achieving automatic ICD encoding for the primary diagnosis of disease. The method was tested on a Chinese dataset with macro-averaged F1 score of 0.862, and the comparative experiments depict that the performance of method based on graph convolutional networks is generally better than that ICD coding models at the text level and node level.