ChatGPT Predicts In-Hospital All-Cause Mortality for Sepsis: In-Context Learning with the Korean Sepsis Alliance...

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Critical Care Medicine

Background:

• Sepsis is a major global cause of death, necessitating accurate outcome prediction for improved patient management.
• Predictive models are crucial for identifying high-risk sepsis patients in intensive care units (ICUs).

Purpose of the Study:

• To evaluate the efficacy of ChatGPT (GPT-3.5-turbo and GPT-4) in predicting in-hospital mortality for sepsis patients.
• To compare ChatGPT's predictive performance against a traditional gradient boosting model (GBM).

Main Methods:

• Utilized data from the Korean Sepsis Alliance (KSA) database (2019-2021) for adult ICU patients.
• Employed structured prompts for in-context learning with varying numbers of patient examples (0-6).
• Assessed predictive accuracy for 7-day and 30-day all-cause mortality using metrics like AUROC.

Main Results:

• GPT-4 showed promising AUROCs for 7-day mortality (0.70-0.83), outperforming GPT-3.5 (0.51-0.70) but comparable to GBM (0.79).
• For 30-day mortality, GPT-4's AUROCs ranged from 0.51-0.59, while GBM achieved 0.76.
• Zero-shot GPT-4 predictions for 30-day mortality yielded AUROCs of 0.60s to 0.75.

Conclusions:

• GPT-4 shows potential for predicting short-term sepsis mortality.
• Performance varied based on evaluation metrics and prediction timeframe.
• Further research is needed to optimize AI models for sepsis outcome prediction.