Multisource representation learning for pediatric knowledge extraction from electronic health records.

Area of Science:

• Biomedical Informatics
• Pediatric Research
• Data Science

Background:

• Electronic Health Record (EHR) systems are crucial for pediatric research but often lack data density.
• Existing EHR embeddings are not optimized for the unique characteristics of pediatric patient data.
• This limits accurate knowledge extraction and patient profiling in pediatric populations.

Purpose of the Study:

• To develop an advanced transfer learning approach for pediatric EHR data.
• To enhance knowledge extraction and relation detection specifically for pediatric contexts.
• To improve patient profiling and identification of pediatric cohorts, such as those with pulmonary hypertension.

Main Methods:

• Introduced MUltisource Graph Synthesis (MUGS), a transfer learning technique.
• Integrated graphical EHR data from pediatric and general sources with medical ontologies.
• Developed adaptive embeddings capturing system homogeneity and heterogeneity for refined EHR feature engineering.

Main Results:

• MUGS embeddings demonstrated superior performance in EHR feature engineering and patient profiling.
• Effectively identified pediatric patients with specific profiles, notably pulmonary hypertension.
• Outperformed benchmark methods, showing resistance to negative transfer in multiple applications.

Conclusions:

• MUGS significantly advances evidence-based pediatric research by providing more accurate EHR data insights.
• The approach effectively addresses the limitations of existing methods for pediatric EHR analysis.
• Enables more nuanced patient profiling and cohort identification in pediatric medicine.