Core Concepts
Pre-trained clinical LLMs offer improved postoperative risk predictions, emphasizing the importance of fine-tuning and foundational modeling for optimal performance in perioperative care.
Abstract
Background:
Postoperative risk predictions are crucial for effective perioperative care management.
Clinical large language models (LLMs) were evaluated for predicting postoperative risks using various training strategies.
Methods:
Utilized 84,875 records from Barnes Jewish Hospital (BJH) and replicated on Beth Israel Deaconess’s MIMIC dataset.
Three domain adaptation and fine-tuning strategies were implemented for BioGPT, ClinicalBERT, and BioClinicalBERT.
Results:
Pre-trained LLMs outperformed traditional word embeddings.
Fine-tuning strategies improved performance, with foundational modeling showing the best results.
Conclusions:
LLMs can enhance postoperative risk predictions in clinical care.
Incorporating labels boosts performance, especially in foundational models.
Data Preprocessing:
Clinical texts were extracted from preoperative and anesthetic records for analysis.
Outcome Variables:
Selected outcomes included 30-day mortality, pulmonary embolism, and pneumonia.
Models Used:
Employed BERT and GPT-based LLMs for postoperative prediction.
ML Predictors and Model Evaluation:
Used XGBoost as the predictor for perioperative outcomes.
Statistical Analysis:
Experiments conducted with 5-fold cross-validation on two datasets.
Discussion:
Demonstrated the potential of LLMs in perioperative care, highlighting the need for further research and data integration.
Stats
Pre-trained LLMs outperformed traditional word embeddings.
Absolute maximal gains of 38.3% for AUROC and 14% for AUPRC were observed.
Incorporating labels improved performance in fine-tuning strategies.
Quotes
"LLMs offer opportunities for postoperative risk predictions in unforeseen data."
"Incorporating labels can boost performance, emphasizing the importance of transferring pre-trained knowledge."