Affective language processing for mental healthcare: A survey of LLM-based approaches across the care continuum

Mental healthcare is fundamentally mediated by affective phenomena expressed through language, including emotional disclosure, shifting mood states, cognitive appraisals, and the relational dynamics of therapeutic alliance. Large language models (LLMs) offer new opportunities to model these signals across mental health workflows, but their clinical value depends on whether they can reliably recognize, track, and respond to affect in context. This survey reviews LLM-based approaches to mental healthcare through an affective computing lens, organizing the literature around emotional expression and affect recognition, empathy and therapeutic alliance, affective dialogue management, longitudinal user-state tracking, and affect-aware safety. We situate these capabilities along a language-mediated care continuum spanning screening and triage, assessment, diagnosis and risk reasoning, and intervention and decision support, linked by iterative follow-up and affective feedback over time. Covering studies published between 2019 and October 2025, we trace the shift from scripted and task-specific NLP systems to LLM-based approaches with stronger contextual understanding and more adaptive response generation, and introduce a three-level taxonomy connecting methodological foundations, service-level applications, and assurance requirements. Across six major research lines, we synthesize current evidence, identify persistent gaps in evaluation and deployment, and outline directions for clinically grounded and patient-centered mental health systems.