Unsupervised Extraction of Diagnosis Codes from EMRs Using Knowledge-Based and Extractive Text Summarization...

Area of Science:

• Medical Informatics
• Natural Language Processing
• Health Information Management

Background:

• Diagnosis codes (ICD-9-CM) are crucial for medical billing, quality control, and research.
• Manual extraction of these codes from electronic medical records (EMRs) by human coders is time-consuming and resource-intensive.
• Existing automated methods often focus on limited document types within EMRs.

Purpose of the Study:

• To propose and evaluate an unsupervised ensemble approach for automatically extracting ICD-9-CM diagnosis codes from EMR narratives.
• To assess the effectiveness of combining Named Entity Recognition (NER), graph-based concept mapping, and extractive text summarization.
• To demonstrate the potential of unsupervised methods in scenarios with limited training data.

Main Methods:

• An unsupervised ensemble model was developed integrating Named Entity Recognition (NER).
• Graph-based concept mapping was employed to represent medical concepts.
• Extractive text summarization techniques were utilized to refine code extraction.
• The model was tested on EMRs from 1000 inpatient visits.

Main Results:

• The unsupervised ensemble approach achieved an example-based average recall of 0.42 and an average precision of 0.47.
• Compared to a baseline using only NER, the graph-based approach improved recall by 12%.
• Extractive text summarization improved precision by 7% over the NER baseline.

Conclusions:

• Unsupervised methods show promise for extracting diagnosis codes from complex EMR narratives, even with long-range dependencies.
• The proposed ensemble approach offers a viable solution for automated ICD-9-CM code extraction, particularly when large training datasets are unavailable.
• This work highlights the potential to enhance efficiency in medical coding and data utilization.