Depression is widely recognized as a major contributor to global disability and a significant factor in the emergence of suicidal tendencies. On social networks, individuals openly share their thoughts and emotions through posts, comments, and other forms of communication. The use of Artificial Intelligence, particularly Machine Learning methods, holds great potential for analyzing this data. However, it is imperative to exercise caution in the application of these methods to avoid biases and overfitting, two problems that could compromise the quality of Machine learning models. In this paper, we present a framework for detecting signs of depression among users of the X social network. This framework is based on four phases aimed at minimizing both biases and overfitting, resulting in models that generalize well to new data, thereby enhancing their applicability by healthcare professionals and patients. To validate our framework, we present the results of three detailed experiments using nine Machine Learning algorithms.