Clinical Text Data in Machine Learning: Systematic Review.

Area of Science:

• Computational linguistics
• Medical informatics
• Machine learning

Background:

• Clinical narratives are vital for healthcare communication and decision-making.
• Natural Language Processing (NLP) unlocks valuable insights from clinical text.
• Machine learning accelerates NLP tool development using large text datasets.

Purpose of the Study:

• To systematically review text data properties for training clinical NLP machine learning models.
• To investigate machine learning-supported NLP tasks and their clinical applications.

Main Methods:

• Systematic review following established guidelines.
• Literature search of MEDLINE via PubMed in August 2018.
• Analysis of 110 studies focusing on data properties, NLP tasks, and clinical applications.

Main Results:

• Most machine learning models used small datasets (hundreds/thousands of documents) due to annotation bottlenecks.
• Active learning and distant supervision were explored to mitigate annotation challenges.
• Text classification for phenotyping, prognosis, and surveillance was a common application.

Conclusions:

• The data annotation bottleneck is a key challenge for clinical NLP.
• Active learning and distant supervision offer solutions for reducing annotation effort.
• Future research should investigate data augmentation, transfer learning, and unsupervised learning.