This research formally outlines the research protocol we used using systematic review methods to evaluate and identify fine-tuned large language models (LLMs) from 2024 and 2025 implementations for mental health use cases involving emotion recognition, depression classification, stress tracking, and counseling dialogue. This study reporting associated quantitative metrics (i.e., accuracy, F1-score). The principal variables to analyze included model architecture, fine-tuning methods, modalities, datasets, and quantitative results. Lexicon-enhanced models like nBERT and domain-specific transformers like BERT-BiLSTM achieved high F1-scores (89%). Although multimodal models improved accuracy in empathy detection, they required complex inputs. Prompt-based models enabled reasoning in empathy detection but resulted in overall lower accuracy. Traditional ML models were genuinely not scalable. Lastly, a meta-analysis suggested that lexicon-guided or domain-adaptive tuning provided a net F1 gain of 12–15% F1-score. However, the lack of consistent evaluation, and limited assessments of human think-aloud processes caused continued limitations in generalizability across tasks. We recommend hybrid fine-tuning approaches integrating lexicon support, personalization, and privacy protections. Finally, we propose RegulaMind, a conceptual framework that combines prompt tuning, emotion lexicons, federated learning, and risk-aware output aligned with the EU AI Act and GDPR.