Clinical risk prediction using language models: benefits and considerations.

Area of Science:

• Artificial Intelligence in Medicine
• Health Informatics
• Clinical Decision Support

Background:

• Electronic Health Records (EHRs) are increasingly used for clinical risk prediction.
• Limited task-specific EHR data often hinders standard machine learning (ML) model performance.
• Leveraging supplementary domain knowledge is crucial for improving EHR-based predictive models.

Purpose of the Study:

• To investigate the potential of Language Models (LMs) to enhance EHR data for clinical risk prediction.
• To develop and evaluate novel LM-based methods for risk prediction using textual EHR data.
• To compare LM-based approaches against traditional methods across diverse datasets.

Main Methods:

• Proposed two novel LM-based methods: LLaMA2-EHR and Sent-e-Med.
• Utilized textual descriptions within structured EHRs for predicting future diagnoses.
• Conducted comparative experiments across 6 methods and 3 risk prediction tasks, varying data types and sizes.

Main Results:

• LM-based representation of structured EHR data, including diagnostic histories, significantly improved risk prediction performance.
• Achieved enhanced performance metrics, including Area Under the Receiver Operating Characteristic (ROC) Curve and Precision-Recall (PR) Curve.
• Demonstrated benefits such as few-shot learning, handling of novel medical concepts, and adaptability to different medical vocabularies.

Conclusions:

• Language Models (LMs) possess extensive embedded knowledge, proving valuable for EHR analysis in risk prediction.
• LM application requires careful consideration due to persistent safety concerns.
• LM-based methods offer a promising avenue for advancing clinical risk prediction using EHR data.