Exploring Named Entity Recognition Potential and the Value of Tailored Natural Language Processing Pipelines for...

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Biomedical Data Science

Background:

• Clinical notes contain unstructured patient data, posing challenges for analysis due to medical jargon and ambiguity.
• Real-time data extraction for clinical decision support tools is complicated by this unstructured nature.
• Named Entity Recognition (NER) is crucial for identifying key patient information within clinical text.

Purpose of the Study:

• To evaluate the data curation, technology, and workflow of a named entity recognition (NER) pipeline for clinical notes.
• To assess the performance of NER models and an NER assertion model in identifying and classifying clinical entities.
• To improve the accuracy of clinical decision support tools through enhanced natural language processing.

Main Methods:

• Utilized Spark NLP clinical pretrained NER models on 138,250 clinical notes from 5000 patients.
• Collected progress care, radiology, and pathology notes, measuring various performance metrics.
• Evaluated NER precision and assertion model precision/recall against expert annotations.

Main Results:

• Achieved excellent NER precision, peaking at 0.989 for procedures, and assertion model accuracy of 0.889.
• Identified long-tail distributions in note characteristics and entity density.
• Progress care notes demonstrated significantly higher entity density compared to radiology and pathology notes.

Conclusions:

• Specialized NER models and pipeline setups are essential for different clinical note types.
• Further research should include diverse clinical notes like discharge summaries and psychiatric evaluations.
• Tailoring NLP approaches enhances performance across varied clinical scenarios.