Drawing on psychological theory, we created a new approach to classify negative sentiment tweets and presented a subset of unclassified tweets to humans for categorization. With these results, a tweet classification distribution was built to visualize how the tweets can fit in different categories. As a final step, we used unsupervised learning to help in the understanding of this new classification, understanding and validating the human factor. The approach developed through visualization, classification and clustering of data could be an important base to measure the efficiency of a machine classifier with psychological diagnostic criteria as the base . Nonetheless, this proposed system used to identify red flags in at risk population for further intervention, due the need to be validated through therapy with an expert.