BEHRT: Transformer for Electronic Health Records

Area of Science:

• Artificial Intelligence in Medicine
• Computational Health Informatics
• Machine Learning for Healthcare

Background:

• Most diagnoses occur after symptom onset, delaying early intervention.
• Precision healthcare requires earlier disease detection for better patient outcomes.
• Machine learning offers potential for proactive disease identification.

Purpose of the Study:

• To introduce BEHRT, a deep neural network for electronic health records (EHR).
• To predict the likelihood of 301 future medical conditions simultaneously.
• To enhance early disease detection and personalized healthcare.

Main Methods:

• Developed BEHRT, a deep neural sequence transduction model.
• Trained and evaluated the model on EHR data from nearly 1.6 million individuals.
• Compared BEHRT's performance against state-of-the-art deep EHR models.

Main Results:

• BEHRT demonstrated significant improvements in average precision scores (8.0-13.2%) over existing models.
• The model achieved high accuracy in predicting 301 future conditions.
• BEHRT showed scalability and enabled personalized interpretation of predictions.

Conclusions:

• BEHRT represents a significant advancement in deep learning for EHR analysis.
• The model enhances early disease prediction accuracy and facilitates personalized medicine.
• BEHRT's architecture supports incorporating heterogeneous data for improved predictive power and offers transferable representations for future research.