Language models are an effective representation learning technique for electronic health record data.

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Clinical Prediction Modeling

Background:

• Electronic health records (EHRs) are widely adopted, enabling machine learning for clinical outcome prediction.
• Training clinical prediction models is often limited by the availability of patient records.

Purpose of the Study:

• To investigate if natural language processing (NLP)-inspired patient representation schemes can enhance clinical prediction model accuracy.
• To assess the impact of these schemes on model performance, particularly with limited training data.

Main Methods:

• Developed patient representation schemes using NLP techniques.
• Applied these schemes to train clinical prediction models on EHR data.
• Evaluated model performance using AUROC (Area Under the Receiver Operating Characteristic curve) on five prediction tasks.

Main Results:

• Patient representation schemes improved AUROC by a mean of 3.5% compared to standard methods.
• The average improvement in AUROC increased to 19% when training data was limited.
• Demonstrated effective information transfer from the general patient population to specific prediction tasks.

Conclusions:

• NLP-inspired patient representation is a viable strategy to boost clinical prediction model accuracy.
• This method is particularly effective in scenarios with scarce patient records for model training.
• Enhances the utility of EHR data for developing robust clinical prediction tools.