Medical records are text documents recording diagnoses, symptoms, examinations, etc. They are accompanied by ICD codes (International Classification of Diseases). ICD is the bedrock for health statistics, which maps human condition, injury, disease etc. to codes. It has enormous financial importance from public health investment to health insurance billing. However, assigning codes to medical records normally needs a lot of human labour and is error-prone due to its complexity. We present a 3-layer attentional convolutional network based on the hierarchy structure of ICD code that predicts ICD codes from medical records automatically. The method shows high performance, with Hit@1 of 0.6969, and Hit@5 of 0.8903, which is better than state-of-the-art method.