Natural Language Processing for EHR-Based Computational Phenotyping.

Area of Science:

• Biomedical Informatics
• Computational Linguistics

Background:

• Electronic Health Records (EHRs) contain vast amounts of unstructured clinical text.
• Extracting meaningful information from EHRs is crucial for advancing medical research and patient care.
• Computational phenotyping aims to identify patient subgroups based on specific characteristics.

Purpose of the Study:

• To review recent advancements in Natural Language Processing (NLP) for computational phenotyping using EHRs.
• To highlight the applications, methods, and challenges in this rapidly evolving field.

Main Methods:

• Survey of NLP techniques applied to EHR data for phenotyping.
• Analysis of keyword search, rule-based systems, supervised, deep learning, and unsupervised learning approaches.
• Discussion on the integration of heterogeneous data sources.

Main Results:

• Keyword and rule-based systems show good performance but require significant manual effort.
• Supervised machine learning models are effective for pattern acquisition.
• Deep learning excels in performance, while unsupervised learning aids in novel phenotype discovery.
• Integrating diverse data modalities improves model performance.

Conclusions:

• NLP is a powerful tool for computational phenotyping from EHRs.
• Machine learning, particularly deep and unsupervised learning, offers scalable and effective solutions.
• Future work should focus on model interpretability, generalizability, and characterizing feature relations in clinical narratives.