Temperature Effects on Large Language Model-Generated Clinical Reports in Critical Care: A Pilot Post Hoc Study

Background: Large language models (LLMs) can generate case-specific clinical guidance informed by SHapley Additive exPlanations (SHAP)-driven intensive care unit (ICU) mortality risk. Temperature is a key adjustable setting that influences output variability and determinism. We evaluated whether adjusting temperature affects the quality and interpretability of GPT-4o clinical reports, and whether similar effects appear in the smaller GPT-4o-mini. Methods: Thirty ICU cases were selected from the Medical Information Mart for Intensive Care IV (MIMIC-IV, version 3.1) database, equally distributed across low, moderate, and high mortality risk groups. SHAP values and predicted risks were provided to GPT-4o using three prompting strategies (P1, P2, and P3) at three temperature settings (0.2, 0.6, and 1.0). We evaluated report quality using the IMPACT (Integration, Mastery, Precision, Applicability, Comprehensiveness, and Timeliness) framework and noted differences across report sections. We also tested GPT-4o-mini under the same conditions for comparison. Results: Only the multi-turn prompting strategy (P3) exhibited improved report quality at higher temperatures, with especially noticeable gains in sections such as Interpretation of Risk Factors. No significant temperature-related changes were observed under P1 or P2 in GPT-4o. In GPT-4o-mini, report quality remained consistent across all temperature settings, demonstrating that the beneficial effects of temperature modulation are evident only in higher-capacity models with sufficient reasoning capability. Conclusion: Temperature tuning in high-capacity LLMs like GPT-4o enhances the richness and precision of clinical reports without sacrificing clarity, offering a pragmatic means to improve support for critical care documentation.