Prediction of Postoperative Vomiting Within 24 Hours Using Machine Learning With Large Language Model-Enhanced...

Area of Science:

• Anesthesiology and Perioperative Medicine
• Artificial Intelligence in Healthcare
• Clinical Informatics

Background:

• Postoperative nausea and vomiting (PONV) are frequent complications following anesthesia.
• Vomiting is a distinct, measurable endpoint crucial for clinical assessment.

Purpose of the Study:

• To develop and validate predictive models for postoperative vomiting within 24 hours.
• To integrate structured perioperative data and unstructured clinical text using a novel framework with Large Language Models (LLMs).
• To separate feature construction from interpretability using LLMs for enhanced clinical understanding.

Main Methods:

• Analysis of 33,460 anesthesia records (2019-2022) with two prediction tasks: preoperative and end-of-surgery.
• Machine learning algorithms (Logistic Regression, XGBoost, LightGBM) for structured data; LLMs for deterministic normalization of unstructured text.
• Performance evaluation using AUC, PR-AUC, calibration, and decision curve analysis; LLM-based QA Chain for post hoc interpretability.

Main Results:

• LightGBM achieved an AUC of 0.729 preoperatively and 0.735 at end-of-surgery, outperforming the Apfel score (AUC 0.610).
• Negative predictive value exceeded 0.95 across models at the Youden-optimal threshold.
• LLM-generated explanations enhanced interpretability but did not significantly improve predictive performance.

Conclusions:

• Machine learning models effectively predict postoperative vomiting using perioperative data.
• LLMs can be integrated reproducibly for deterministic normalization and post hoc reasoning, enhancing model interpretability.
• Further clinician-centered validation and external multicenter studies are necessary to confirm clinical utility and applicability.