Non-suicidal self-injury, often referred to as self-injury, is the act of deliberately
harming one’s own body, such as cutting or burning oneself. It is not usually
intended as a suicide attempt. This paper presents a system for detecting signs of
non-suicidal self-injury, based on language analysis, on an annotated set of medical
reports obtained from the psychiatric service of a public hospital in Madrid. Both
explainability and accuracy in predicting positive cases are the two main objecti-
ves of this work. In order to achieve this goal, two supervised systems of different
natures have been developed. On the one hand, a process of extraction of different
features focused on the world of self-injury itself has been carried out using natural
language processing techniques to subsequently feed a traditional classifier. On the
other hand, a deep learning system based on several layers of convolutional neural
networks, due to its high performance in text classification tasks. The result are two
supervised systems with high performance, where we highlight the system based on
a traditional classifier due to its better prediction of positive classes and the greater
ease to explain its results to health professionals.