The high prevalence of mental illnesses in adolescents makes it crucial to design innovative ways to identify and intervene at an early stage. This paper presents a new framework based on Bidirectional Encoder Representations from Transformers to detect potential mental health risks of teenagers from their online social network activities. The framework captures the subtle linguistic nuances used by teenagers experiencing mental health struggles. This is achieved by fine-tuning a BERT-based sequence classification model on a large dataset of social media posts, categorized across multiple mental health conditions, with the intent of proactive identification of at-risk teenagers. It presents an evaluation of this model, developed using a stratified k-fold cross-validation approach, scoring an average accuracy of 81.42%, precision of 81.98%, recall of 81.49%, and F1-score of 81.42%. The research is thus important for the field of artificial intelligence in mental health screening and points toward strong potential for natural language processing to help better refine early intervention strategies for adolescent mental health.