Information about causes of death are collected by filling death certificates according to a standard defined by WHO. Conditions are expressed by means of the International Statistical Classification of Diseases and Related Health Problems, revision 10. Starting from such information, the so-called Underlying Cause of Death is selected with statistical and epidemiological aims. This task is usually carried out with the support of a system that implements the rules defined by WHO. However. such systems leave the selection of a fraction of certificates up to 31.5% of the total to the interpretation of the human expert, causing a significant burden for the Institutes in charge of coding. Machine learning has been used to both code free text certificates to ICD-10, and recently also to identify the underlying cause of death. The present paper describes a method for selecting the underlying cause of death from death certificates, based on the use of categorical embeddings and convolutional neural network. The proposed method achieves 98.44% accuracy, which is currently state-of-the-art performance.