Bidirectional Representation Learning From Transformers Using Multimodal Electronic Health Record Data to Predict...

Area of Science:

• Artificial Intelligence
• Biomedical Informatics
• Machine Learning

Background:

• Electronic Health Records (EHR) offer rich data for predictive modeling.
• Improving the performance and interpretability of machine learning models in healthcare is crucial.
• Predicting chronic diseases like depression from EHR data remains a challenge.

Purpose of the Study:

• To develop and evaluate a temporal deep learning model for depression prediction using EHR data.
• To enhance the interpretability of machine learning models in clinical decision support.
• To aggregate and process heterogeneous EHR data for chronic disease prediction.

Main Methods:

• Utilized a transformer-based deep learning architecture for bidirectional representation learning on EHR sequences.
• Integrated five heterogeneous, high-dimensional data sources from EHRs.
• Applied pretraining and fine-tuning techniques on EHR data.
• Employed self-attention mechanisms to interpret model decisions.

Main Results:

• Achieved a precision-recall area under the curve (PRAUC) of 0.76 for depression prediction, outperforming baseline models.
• Demonstrated significant improvement in depression prediction accuracy compared to state-of-the-art methods.
• Self-attention weights quantitatively revealed relationships between EHR codes, enhancing model interpretability.

Conclusions:

• The temporal deep learning model effectively predicts depression using heterogeneous EHR data with high accuracy and interpretability.
• This approach shows promise for developing clinical decision support systems for chronic disease screening and early detection.
• The model's ability to interpret EHR code relationships facilitates understanding of prediction drivers.