CPLLM: Clinical prediction with large language models.

Area of Science:

• Artificial Intelligence
• Clinical Informatics
• Medical Data Science

Background:

• Predicting clinical disease and hospital readmission is crucial for patient care and healthcare management.
• Existing methods for analyzing electronic health records (EHR) often require specialized pre-training or struggle with temporal data.
• Large Language Models (LLMs) show promise in processing complex data but require adaptation for clinical prediction tasks.

Purpose of the Study:

• To introduce Clinical Prediction with Large Language Models (CPLLM), a novel method for disease diagnosis and hospital readmission prediction.
• To evaluate CPLLM's performance against established baseline models, including state-of-the-art Med-BERT.
• To demonstrate CPLLM's effectiveness without requiring pre-training on specific medical datasets.

Main Methods:

• Fine-tuning a pre-trained Large Language Model (LLM) using quantization and prompt-based techniques.
• Utilizing historical patient medical records for diagnostic predictions (next visit or subsequent diagnosis).
• Evaluating CPLLM for hospital readmission prediction and comparing against benchmark models.

Main Results:

• CPLLM achieved superior performance compared to all tested baseline models.
• The method demonstrated state-of-the-art results in both disease prediction and hospital readmission prediction, measured by PR-AUC and ROC-AUC metrics.
• CPLLM successfully predicted clinical outcomes without requiring pre-training on medical data.

Conclusions:

• CPLLM offers a powerful and effective approach for clinical prediction tasks, including disease diagnosis and hospital readmission.
• The method's ability to perform without medical pre-training simplifies implementation and integration into clinical workflows.
• CPLLM has the potential to significantly aid healthcare providers in planning patient care and interventions.