KMTLabeler: An Interactive Knowledge-Assisted Labeling Tool for Medical Text Classification.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Accurate medical text labeling is vital for research but is often time-consuming and requires domain expertise.
• Traditional rule-based methods lack adaptability, while current machine learning (ML) approaches can be challenging for non-technical users and may not fully integrate expert input.
• Existing semi-automated methods like data programming may not offer continuous refinement of results.

Purpose of the Study:

• To develop an efficient collaborative human-ML teaming workflow for medical text labeling.
• To introduce an innovative neural network model (embedding network) for generating task-specific medical text embeddings incorporating expert insights.
• To integrate these components into a visual analytics tool (KMTLabeler) for enhanced labeling efficiency.

Main Methods:

• A collaborative workflow integrating visual cluster analysis and active learning.
• Development of an embedding network model that incorporates expert insights for medical text representation.
• Implementation of KMTLabeler, a visual analytics tool with coordinated multi-level views and interactions.

Main Results:

• Demonstrated effectiveness of KMTLabeler in creating an efficient medical text labeling environment.
• Successful integration of human expertise and ML through visual analytics and active learning.
• Generation of task-specific embeddings for medical texts using the novel embedding network.

Conclusions:

• KMTLabeler significantly improves the efficiency of medical text labeling for domain experts.
• The proposed human-ML teaming workflow and embedding network offer a novel approach to medical data annotation.
• The findings support the use of KMTLabeler in medical research for streamlined and accurate text classification.