Clinical knowledge extraction via sparse embedding regression (KESER) with multi-center large scale electronic health...

Area of Science:

• Computational biology
• Health informatics
• Translational research

Background:

• Electronic Health Record (EHR) systems offer vast potential for translational research.
• Identifying phenotype-related codes is challenging due to the large volume of available codes.
• Traditional data mining requires patient-level data, limiting cross-institutional data sharing.

Purpose of the Study:

• To demonstrate the efficacy of multi-center, large-scale code embeddings for identifying disease-related features.
• To develop and evaluate a novel method for feature selection and knowledge extraction from EHR data.
• To create an integrated clinical knowledge map for enhanced multi-institutional analysis.

Main Methods:

• Constructed large-scale code embeddings from EHR data across two medical centers.
• Developed Knowledge Extraction via Sparse Embedding Regression (KESER) for feature selection.
• Performed integrative network analysis and evaluated KESER's performance on eight diseases.

Main Results:

• KESER identified comprehensive features, comparable to expert-curated lists.
• Features selected by KESER achieved performance similar to manual or patient-level data methods.
• An integrated knowledge map improved accuracy in identifying disease-disease and disease-drug relationships compared to single-institution data.

Conclusions:

• Analysis of code embeddings via KESER effectively reveals clinical knowledge and concept relatedness.
• KESER facilitates multi-center EHR studies by eliminating the need for patient-level data sharing.
• This approach significantly advances the utilization of EHR data for large-scale research.