Disease Concept-Embedding Based on the Self-Supervised Method for Medical Information Extraction from Electronic...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Data Science

Background:

• Electronic Health Records (EHRs) contain valuable patient data but often require organization for effective clinical use.
• Limited patient information in some cases necessitates robust tools for treatment decision support.
• Natural Language Processing (NLP) and deep learning can transform EHR data into actionable medical knowledge.

Purpose of the Study:

• To develop a model for extracting concept embeddings from EHRs for disease pattern retrieval.
• To enable further classification tasks using these extracted embeddings.
• To enhance clinical decision-making through improved access to organized medical information.

Main Methods:

• Utilized a Transformer-based model incorporating Bidirectional Encoder Representations from Transformers (BERT) for feature extraction from EHRs.
• Employed Deep InfoMax (DIM) and Simple Contrastive Learning of Visual Representations (SimCLR) for unsupervised disease concept embedding.
• Pretrained the EDisease model and finetuned it for critical care outcome prediction, evaluating performance using Area Under the Receiver Operating Characteristic (AUROC).

Main Results:

• The EDisease model achieved a high AUROC of 0.876 for outcome prediction.
• Ablation studies indicated that pretraining significantly improved prediction performance compared to models without pretraining.
• Performance decreased with smaller datasets or fewer unsupervised pretraining methods, highlighting the importance of the proposed approach.

Conclusions:

• Contrastive learning effectively embeds disease concepts, facilitating disease retrieval and enhancing clinical practice.
• The developed disease concept model serves as a valuable pretrained model for various downstream prediction tasks.
• The findings demonstrate the potential of AI-driven EHR analysis to improve healthcare outcomes.