Eating disorders (EDs) are the mental illnesses with the highest mortality rate, resulting in around 10,200 annual deaths worldwide. However, with early treatment, the probability of dying from an eating disorder can be reduced from 19-20% to 2-3%. Many techniques have been developed to leverage social media to detect whether an individual is suffering or not from this mental health condition. However, many of them are subject to the number of posts uploaded by the user being analysed rather than taking other factors into account. In this work, we present a novel approach for the early detection of eating disorders that maximises the amount of information that can be extracted from a single post. We use a set of features characterising eating disorders related content, the embedding of the texts and the personal information contained in each post to train a feature fusion model able to identify all test sample users at risk, achieving a 100% recall and 0.97% F1 score for the positive class.