An interactive and user-centered computer system to predict physician's disease judgments in discharge summaries

Area of Science:

• Medical Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• Clinical text analysis is crucial for healthcare but often complex.
• Natural Language Processing (NLP) systems aim to automate information extraction from clinical notes.
• User-centered design principles can enhance the usability and effectiveness of NLP tools.

Purpose of the Study:

• To describe a formative NLP system built on user-centered design, simplification, and functional transparency.
• To evaluate the system's performance in classifying diseases within patient discharge summaries.
• To assess the impact of interactive training and a de novo knowledge base on classification accuracy.

Main Methods:

• The NLP system employed interactive, fully supervised learning.
• Rule-based processes and Support Vector Machines (SVMs) were utilized for text classification.
• The system was evaluated in the 2008 i2b2 Shared Task competition against clinician judgment.

Main Results:

• Macro-averaged F-scores for textual and intuitive classification were 0.614 and 0.629, respectively.
• Micro-averaged F-scores reached 0.966 (textual) and 0.954 (intuitive).
• Performance was comparable to the top 10 systems in the competition.

Conclusions:

• Interactive training, a de novo knowledge base, and simplified text mining achieve high performance in health text classification.
• User-centered NLP system design shows promise for clinical applications.
• Further research is required to confirm real-world benefits of user-centered NLP systems.