Clinical Sentiment Analysis by Large Language Models Enhances Prediction of Hepatorenal Syndrome

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Clinical Prediction Models

Background:

• Hepatorenal syndrome-acute kidney injury (HRS-AKI) is a severe complication of decompensated cirrhosis with challenging prediction.
• Traditional predictors for HRS-AKI often lack sufficient accuracy.
• Large language models (LLMs) show promise in natural language processing (NLP) for clinical data.

Purpose of the Study:

• To determine if LLMs can derive clinical sentiment from unstructured text.
• To assess if this sentiment analysis improves HRS-AKI prediction over traditional methods.
• To evaluate the feasibility of using GPT-4o for clinical NLP tasks.

Main Methods:

• A cohort study of 314 adult patients with decompensated cirrhosis and AKI was conducted.
• Microsoft Azure OpenAI GPT-4o was used to derive HRS-AKI sentiment scores and extract clinical terms.
• Logistic regression models compared AUROC with and without LLM-derived features.

Main Results:

• Higher sentiment scores correlated with HRS-AKI diagnosis (OR: 1.52 per 0.1).
• The AUROC using structured data alone was 0.63.
• Incorporating GPT-4o sentiment and NLP terms improved AUROC to 0.79 (P < .01).

Conclusions:

• Clinical texts contain valuable data not easily extracted by standard methods.
• Sentiment analysis and NLP with GPT-4o are feasible for clinical applications.
• LLM-derived features enhance HRS-AKI prediction accuracy compared to structured data alone.