Multi-layer Representation Learning and Its Application to Electronic Health Records

Area of Science:

• Clinical Informatics
• Data Science
• Machine Learning

Background:

• Electronic Health Records (EHRs) contain valuable patient data but present challenges due to temporal visit ordering and random diagnosis code ordering within visits.
• Secondary use of EHRs is crucial for advancing clinical informatics and healthcare development.
• Existing methods struggle to capture the hierarchical and relational information inherent in EHR data.

Purpose of the Study:

• To propose a novel Multi-Layer Representation Learning (MLRL) method for effective patient representation learning from EHRs.
• To address the unique hierarchical structure of EHRs by exploring relationships at both diagnosis code and patient visit levels.
• To enhance the performance of predictive models using learned patient representations.

Main Methods:

• MLRL employs a multi-head attention mechanism to identify connections among diagnosis codes, followed by a linear transformation for vector representation.
• Initial visit vectors are generated by summarizing diagnosis code representations.
• Bidirectional Long Short-Term Memory with self-attention is used to learn weighted visit vectors, aggregated into a final patient representation.

Main Results:

• MLRL achieved a significant improvement in patient mortality prediction performance on real EHR data.
• The method attained an Area Under the Curve (AUC) of approximately 0.915, outperforming baseline methods.
• Learned representations from MLRL demonstrated superior performance and availability across multiple classifiers compared to raw data and other representations.

Conclusions:

• MLRL effectively captures hierarchical and relational information from EHRs, leading to superior patient representation.
• The proposed method offers a significant advancement in leveraging EHR data for clinical informatics applications, particularly in mortality prediction.
• MLRL shows strong potential for improving various downstream machine learning tasks in healthcare.