Data-Driven Information Extraction from Chinese Electronic Medical Records.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Data Science

Background:

• Electronic Medical Records (EMRs) contain unstructured free text narratives.
• Extracting structured information from Chinese EMRs presents unique challenges due to linguistic conventions.
• Existing methods may not adequately capture the temporal and descriptive nuances within medical narratives.

Purpose of the Study:

• To propose a novel data-driven framework for converting unstructured Chinese EMR narratives into structured time-event-description triples.
• To develop a system capable of identifying medical events, their temporal context, and associated descriptions.
• To enhance the utility of Chinese EMRs for clinical research and data analysis.

Main Methods:

• A hybrid approach combining cross-domain medical lexica construction and an unsupervised iterative algorithm for term accrual.
• Development of rules to handle Chinese writing conventions and temporal descriptors.
• Implementation of a Support Vector Machine (SVM) algorithm utilizing Normalized Google Distance (NGD) for event-description correlation.

Main Results:

• The framework achieved an F1-score of 0.896 for medical term recognition.
• 98.5% of medical events were successfully linked to temporal descriptors.
• The end-to-end extraction of time-event-description triples yielded an F1-score of 0.846.
• The NGD SVM demonstrated superior performance in event-description association (F1-score: 0.874).

Conclusions:

• The proposed framework outperforms state-of-the-art supervised learning algorithms in named entity recognition.
• The NGD SVM approach is more effective for event-description association than traditional SVM methods.
• The framework is data-driven, weakly supervised, robust, and adaptable to variations in Chinese medical writing.