A scoping review of self-supervised representation learning for clinical decision making using EHR categorical data

Area of Science:

• Clinical Informatics
• Artificial Intelligence in Medicine
• Data Science

Background:

• Electronic Health Records (EHRs) generate vast amounts of unlabeled categorical data.
• Deep learning, specifically Self-Supervised Representation Learning (SSRL), offers a powerful approach to extract meaningful insights from this data.
• The integration of SSRL into clinical decision-making processes is rapidly evolving.

Purpose of the Study:

• To conduct a scoping review of studies utilizing SSRL for unlabeled categorical EHR data.
• To systematically assess research trends, model families, and applications of SSRL in healthcare.
• To identify limitations and future research opportunities for SSRL in clinical practice.

Main Methods:

• Systematic literature search across PubMed, MEDLINE, Embase, ACM, and Web of Science.
• Inclusion of 46 studies published between January 2019 and April 2024.
• Analysis based on PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines.

Main Results:

• Transformer-based models (43%) dominate SSRL research for EHR data, followed by Autoencoder-based (28%) and Graph Neural Network-based (17%) models.
• Identified trends focus on developing computationally and data-efficient representations for medical tasks.
• Highlighted are practical scenarios for healthcare institutions to adopt or develop SSRL technologies.

Conclusions:

• SSRL shows significant potential for enhancing clinical decision-making through EHR data analysis.
• Further research is needed to overcome limitations in impact assessment and fully realize SSRL's influence on clinical practice.
• Healthcare institutions should explore leveraging SSRL for improved data-driven insights and operational efficiency.