TAPER: Time-Aware Patient EHR Representation.

Area of Science:

• Medical informatics
• Natural Language Processing
• Machine Learning

Background:

• Electronic health records (EHR) contain complex, multi-modal data including clinical notes and medical codes.
• Extracting meaningful insights from irregular EHR data is crucial for improving patient care and outcomes.
• Existing methods struggle to effectively integrate diverse data streams within EHRs.

Purpose of the Study:

• To develop an effective representation learning method for electronic health records.
• To unify disparate data modalities (notes, codes) into a single vector representation.
• To enhance downstream predictive tasks using patient visit data.

Main Methods:

• Utilized transformer networks and the BERT language model for embedding EHR data streams.
• Developed a unified vector representation for patient visit data.
• Applied the model to the publicly available MIMIC-III ICU dataset.

Main Results:

• The proposed model achieved superior performance in predicting patient mortality.
• Demonstrated enhanced generalization capabilities for readmission prediction.
• Showcased effectiveness in predicting length of stay for ICU patients.

Conclusions:

• Transformer networks and BERT provide a powerful approach for EHR representation learning.
• The unified vector representation facilitates improved predictive accuracy on critical healthcare tasks.
• This method offers a robust solution for leveraging complex EHR data for clinical decision support.