Large language models and multimodal AI for mental health: a systematic review of early diagnosis and monitoring

Mental health disorders (e.g., depression, anxiety, post-traumatic stress disorder (PTSD), bipolar disorder) represent a pressing global challenge, and early diagnosis with continuous monitoring is critical for effective intervention. However, traditional diagnostic methods, relying on patient self-reports and clinical interviews, are subjective and often miss subtle early warning signs, a problem compounded by stigma and limited access to care. In response, recent advances in artificial intelligence, particularly large language models (LLMs) and multimodal learning techniques, may support mental-health screening, monitoring, and decision support by enabling scalable, data-driven, and personalized analysis of various behavioral and physiological signals. This review provides a comprehensive integration of current LLM-based and multimodal approaches for early mental health diagnosis and monitoring. We examine key AI frameworks and their architectures, highlighting how they integrate modalities such as text, speech, neuroimaging, and physiological data through various fusion strategies (early, late, hybrid, and cross-attention) to improve predictive accuracy. In addition, we critically discuss the technical, ethical, and clinical challenges, such as model hallucinations, algorithmic bias, interpretability limitations, and privacy concerns, that must be addressed to translate these technologies into reliable clinical tools. Finally, we outline open challenges and future research directions aimed at guiding researchers, clinicians, and developers in advancing responsible, human-centered AI systems for mental health care.