Development and Validation of a Natural Language Processing Algorithm for Extracting Clinical and Pathological...

Area of Science:

• Oncology
• Medical Informatics
• Artificial Intelligence

Background:

• Electronic health records (EHRs) are rich sources of real-world data for breast cancer (BC) research.
• Extracting structured data from unstructured EHRs, particularly pathology reports, is challenging.
• Natural Language Processing (NLP) offers a potential solution for automating data extraction.

Purpose of the Study:

• To develop and evaluate an NLP model for extracting structured data from BC pathology reports.
• To compare the performance of a rule-based NLP algorithm with a Named Entity Recognition (NER)-NLP model.
• To facilitate the creation of comprehensive BC databases for clinical research.

Main Methods:

• Initial development used a rule-based NLP algorithm on 193 BC pathology reports (2012-2016).
• Data extraction variables were compared against manual extraction by a data specialist and an oncologist.
• A subsequent NER-NLP model was trained and validated on an expanded dataset of 513 reports using K-fold cross-validation.

Main Results:

• The rule-based NLP algorithm achieved 82.9% concordance with an oncologist.
• Manual extraction by a data specialist showed 90.8% concordance with an oncologist.
• The NER-NLP model demonstrated high accuracy (97.8%), with perfect performance (F1-score 1.0) for key biomarkers like ER, PR, HER2, and Ki-67.

Conclusions:

• The developed NLP model shows significant potential for accelerating the creation of cancer databases.
• AI-driven data extraction from EHRs can enhance clinical research and support oncology initiatives.
• Further postprocessing aims to organize extracted data into usable tabular formats for research and clinical applications.