Automatic extraction of numerical values from unstructured data in EHRs.

Area of Science:

• Clinical Informatics
• Natural Language Processing
• Health Data Science

Background:

• Electronic Health Records (EHRs) contain rich clinical data, but secondary use for research and decision-making is hindered by data quality issues like missing values, inaccuracies, and information locked in unstructured narrative documents.
• Extracting meaningful information from narrative clinical text is a significant challenge in health informatics.

Purpose of the Study:

• To develop and evaluate a system for detecting and extracting numerical values from unstructured clinical documents.
• To associate extracted numerical values with their corresponding medical concepts (themes) and units to create semantically meaningful data sequences.

Main Methods:

• Utilized a Conditional Random Fields (CRF) supervised categorization model to identify segments including themes, numerical sequences, and units within narrative clinical text.
• Employed a rules-based system to associate these identified segments, constructing semantically coherent data sequences.

Main Results:

• Achieved competitive performance metrics: 0.96 precision, 0.78 recall, and 0.86 F-measure.
• Demonstrated the system's capability to effectively process narrative clinical documents for data extraction.

Conclusions:

• The proposed system offers a viable solution for overcoming challenges in extracting structured data from unstructured clinical notes.
• The successful results indicate potential for application with larger-scale clinical datasets to enhance data accessibility and utility.