Prompt-Sensitive Decision Behavior of Large Language Models in Intensive Care Unit Mortality Prediction for...

Area of Science:

• Clinical prediction modeling
• Artificial intelligence in healthcare
• Medical informatics

Background:

• Large language models (LLMs) are being explored for clinical decision support.
• Reliability of LLM-generated quantitative risk estimates in structured clinical prediction is uncertain.

Purpose of the Study:

• Evaluate LLM predictive performance and decision-making in structured clinical prediction.
• Compare LLM outputs with an outcome-trained machine learning model.

Main Methods:

• Benchmarking study using structured clinical data from intensive care unit patients with intracerebral hemorrhage.
• Compared an extreme gradient boosting model with a general-purpose LLM using various prompting strategies (zero-shot, few-shot, chain-of-thought).
• Evaluated discrimination, classification behavior, and feature importance concordance.

Main Results:

• Outcome-trained model outperformed all LLM approaches in discriminative performance.
• LLMs showed moderate discrimination but significant variability in classification behavior across prompts.
• LLM optimal thresholds were substantially higher (0.74-0.88) than the machine learning model (0.1555).
• Modest concordance between LLM feature prioritization and model-derived feature importance.

Conclusions:

• Inference-only LLM outputs in structured clinical prediction are prompt-sensitive and require cautious interpretation for risk estimation.
• Integrating outcome-trained models with LLM-assisted reasoning may enhance future clinical decision support systems.