Large language models are less effective at clinical prediction tasks than locally trained machine learning models

Area of Science:

• Artificial Intelligence in Healthcare
• Clinical Predictive Modeling
• Electronic Health Records (EHR) Data Analysis

Background:

• Large Language Models (LLMs) are increasingly explored for various applications, including healthcare.
• Traditional Machine Learning (ML) methods are established for clinical prediction using Electronic Health Records (EHRs).
• Evaluating LLMs as potential substitutes for traditional ML in clinical settings is crucial for advancing healthcare technology.

Purpose of the Study:

• To assess the efficacy of current LLMs (GPT-3.5, GPT-4) as clinical predictors compared to traditional ML.
• To investigate factors influencing LLM adoption in clinical prediction: performance, calibration, fairness, and privacy resilience.
• To analyze the impact of data generalization for privacy on model performance.

Main Methods:

• Comparative analysis of GPT-3.5, GPT-4, and gradient-boosting trees (traditional ML) on EHR data from VUMC and MIMIC IV.
• Performance measured by Area Under the Receiver Operating Characteristic (AUROC) and Brier Score for calibration.
• Fairness evaluated using equalized odds and statistical parity across demographic groups.
• Impact of in-context learning and data generalization for privacy on AUROC was assessed.

Main Results:

• Traditional ML demonstrated substantially superior predictive performance (AUROC: 0.847, 0.894) compared to GPT-3.5 (AUROC: 0.537, 0.517) and GPT-4 (AUROC: 0.629, 0.602).
• Traditional ML also exhibited significantly better output probability calibration (Brier Score: 0.134, 0.042) than GPT-3.5 (0.384, 0.06) and GPT-4 (0.251, 0.219).
• GPT-4 showed better fairness across demographic groups but at the expense of reduced model performance.

Conclusions:

• Non-fine-tuned LLMs are currently less effective and robust than locally trained ML for clinical prediction using EHR data.
• Traditional ML models are more robust in generalizing demographic information for privacy protection.
• While current LLMs lag behind traditional ML, ongoing advancements suggest potential for future clinical applications.