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
• * Health Data Science

Background:

• * Electronic Health Records (EHRs) are increasingly prevalent, generating vast amounts of unlabeled categorical data.
• * Deep learning, especially Self-Supervised Representation Learning (SSRL), offers powerful methods for extracting insights from this data.
• * SSRL is crucial for developing computationally and data-efficient representations for medical tasks.

Purpose of the Study:

• * To conduct a scoping review of studies applying SSRL to unlabeled categorical EHR data.
• * To identify research trends, dominant model families, and applications of SSRL in healthcare.
• * To highlight opportunities and limitations for leveraging SSRL in clinical practice.

Main Methods:

• * A systematic scoping review adhering to PRISMA-ScR guidelines.
• * Analysis of 46 studies published between January 2019 and April 2024.
• * Data sourced from major scientific databases: PubMed, MEDLINE, Embase, ACM, and Web of Science.

Main Results:

• * Transformer-based models (43%) were the most prevalent, followed by Autoencoder-based (28%) and Graph Neural Network-based (17%) models.
• * Identified key trends in building efficient data representations for various medical tasks.
• * Highlighted specific scenarios for healthcare institutions to adopt or develop SSRL technologies.

Conclusions:

• * SSRL shows significant promise for enhancing clinical decision-making through efficient EHR data representation.
• * Current limitations exist in assessing the real-world impact of these technologies.
• * Future research should focus on overcoming these limitations and increasing the influence of SSRL on clinical practice.