DOME: Directional medical embedding vectors from Electronic Health Records.

Area of Science:

• Biomedical Informatics
• Machine Learning
• Translational Research

Background:

• Electronic Health Record (EHR) systems offer significant potential for translational research.
• Representation learning and knowledge graphs enhance EHR studies, but often require patient-level data, limiting multi-institutional use.
• Existing scalable methods using summary-level data lack temporal dependency modeling.

Purpose of the Study:

• To introduce a novel algorithm, DirectiOnal Medical Embedding (DOME), for encoding temporal relationships between medical concepts using summary-level EHR data.
• To address the limitations of existing methods by enabling multi-institutional EHR data training and incorporating temporal dynamics.

Main Methods:

• DOME aggregates patient-level EHR data into an asymmetric co-occurrence matrix.
• It computes Positive Pointwise Mutual Information (PPMI) matrices to capture prior and posterior dependencies between medical concepts.
• Joint matrix factorization of PPMI matrices yields semantic and directional context embeddings for each concept.

Main Results:

• DOME improves disease risk prediction, showing a 5.5% relative gain in AUROC for lung cancer prediction.
• It enhances directional drug-disease relationship inference, outperforming state-of-the-art methods with relative AUROC gains of 10.8% and 6.6% for differentiating side effects and indications.
• DOME constructs directional knowledge graphs that reveal disease progression trajectories by distinguishing risk factors from comorbidities.

Conclusions:

• DOME effectively models temporal relationships in EHR data using only summary-level information.
• The algorithm demonstrates significant translational potential in improving disease risk prediction, drug-disease inference, and understanding disease progression.
• DOME provides a valuable tool for leveraging multi-institutional EHR data in biomedical research.