Successful Development of a Natural Language Processing Algorithm for Pancreatic Neoplasms and Associated Histologic...

Area of Science:

• Computational pathology
• Medical informatics

Background:

• Pancreatic pathologies encompass a wide range of benign and malignant conditions.
• Accurate interpretation of histologic features is crucial for diagnosis and treatment.
• Electronic health records contain unstructured text vital for research and clinical applications.

Purpose of the Study:

• To develop and evaluate a natural language processing (NLP) algorithm for analyzing pancreatic pathology reports.
• To improve the efficiency of electronic health record coding for pancreatic diseases.
• To facilitate the creation and curation of large-scale pancreatic disease databases.

Main Methods:

• A dataset of over 14,000 text-based pancreatic cytopathologic reports was collected.
• Reports included diagnoses such as pancreatic cancer, ductal adenocarcinoma, and neuroendocrine tumors.
• A convolutional neural network was trained on 1252 reports for pathology prediction, with 80/20 split for training/development and a separate test set.

Main Results:

• The NLP algorithm achieved high accuracy and F1 scores, ranging from 95% to 98% on the test set.
• Learning curves demonstrated improved performance with increased training data volume.
• Some queries exhibited high index performance, indicating robust algorithm capabilities.

Conclusions:

• Natural language processing (NLP) algorithms are effective for analyzing pancreatic pathologies.
• Factors such as increased training data, distinct terminology, and consistent text structure enhance NLP performance.
• The developed algorithm shows promise for clinical applications in pathology report analysis.