As mental health disorders continue to rise globally and access to specialised care remains a persistent challenge, AI-powered conversational agents have emerged as a credible avenue for early screening. This systematic review, conducted in ac cordance with PRISMA guidelines and drawing on databases including PubMed, Springer, and IEEE Xplore, critically assesses the diagnostic validity and clinical e ectiveness of a range of NLP models , from classical architectures such as CNN and SBERT-CNN to Transformer-based models, most notably RoBERTa. Studies reporting standardised clinical performance metrics and user acceptability data were retained; those relying on insu cient sample sizes or purely static applications were excluded. Findings indicate that classical models are constrained in their semantic modelling capacity, whereas Transformer-based architectures, RoBERTa in particu lar, demonstrate richer contextual representation, enabling more nuanced detection of subtle linguistic cues. Despite superior performance, RoBERTa still poses notable challenges: high computational cost, dependence on annotated training corpora, and susceptibility to linguistic bias. In conclusion, RoBERTa stands out as the most ef fective model for early mental disorder screening; however, its clinical integration calls for lighter architectures, improved interpretability, and robust ethical frame works grounded in meaningful human oversight.