Abstract
Large language models (LLMs) are being deployed in clinical settings despite an underdeveloped evidence base regarding their real-world effectiveness. This study employed systematic evidence mapping to characterize outcome measures used in published studies and registered clinical trials (Jan 2022-Jun 2025) evaluating LLM performance. Analysis of 55 included studies revealed a predominance of human-AI collaborative designs (65.5%) for decision support and symptom management. LLM-only interventions focused on functional performance and operational or process impact outcomes (e.g., accuracy and time saving), whereas LLM-assisted interventions showed positive clinical effects, particularly in psychological health endpoints. Critical evidence gaps persist: diagnostic accuracy in randomized trials was notably lower and more variable (range 0.65-0.88) compared to non-randomized studies (typically ≥ 0.80); clinical efficiency impacts were inconsistent, and reporting quality was suboptimal (78.8% mean CONSORT-AI adherence), with critical omissions in handling data quality and performance errors. These findings indicate a heterogeneous and insufficient evidence landscape, necessitating standardized core outcome sets, mandatory use of specialized reporting guidelines, and robust clinical trials to ensure the safe integration of LLMs.