Natural Language Processing to extract SNOMED-CT codes from pathological reports.

Area of Science:

• Digital Pathology
• Medical Informatics
• Artificial Intelligence in Healthcare

Background:

• Standardized structured reports (SSR) and terminologies like SNOMED-CT are crucial for data retrieval and analysis in pathology.
• The prevalence of narrative reports hinders large-scale studies and collaboration, necessitating automated labeling solutions.

Purpose of the Study:

• To develop and evaluate natural language processing (NLP) methods for automatically associating SNOMED-CT codes with digital pathology reports.
• To address the challenge of organizing unstructured narrative reports in pathology archives.

Main Methods:

• Two NLP-based automatic coding systems, Support Vector Machine (SVM) and Long Short-Term Memory (LSTM), were trained and applied to narrative pathology reports.
• Explainability features were integrated to identify important terms and optimize model performance.

Main Results:

• Both SVM and LSTM models achieved good performance metrics (accuracy, precision, recall, F1 score) on 1163 cases.
• The SVM model exhibited slightly superior performance compared to the LSTM model.
• Explainability facilitated model fine-tuning by highlighting key terms and word groups.

Conclusions:

• AI tools, specifically NLP, can automate SNOMED-CT labeling of pathology archives.
• This approach offers a retrospective solution for the lack of organization in narrative pathology reports.
• Automated coding enhances data accessibility and supports future research and collaboration.