LLMental: Classification of Mental Disorders with Large Language Models

The increasing number of mental disorders is a severe problem in the modern world and can even lead to suicide if left untreated. In the age of digitalization, we move part of our lives to social media, where we share both the good and the bad moments. This allows for the early detection of mental disorders (such as depression, excessive stress, or social phobia) of which the user may even be unaware. We propose to modify large language models, such as PHI-2, Mistral, Flan-T5, or LLaMA 2, to classify mental disorders and to add appropriate layers. This gives a better prediction performance than zero-shot/few-shot for LLMs and classification by BERT-based models. Using such an architecture makes it possible to return a label rather than text, thus allowing the output of the LLM model to be freely modified.