Leveraging Hybrid Natural Language Processing Techniques for Large-Scale Pulmonary Embolism Identification.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Health Data Analysis

Background:

• Pulmonary embolism (PE) diagnosis and management remain challenging despite medical advancements.
• Large-scale electronic health record analysis is crucial for understanding PE pathophysiology and optimizing treatment.
• Hybrid natural language processing (NLP) offers a promising methodology for this analysis.

Purpose of the Study:

• To develop and validate a hybrid NLP pipeline for accurate PE case identification.
• The pipeline combines machine learning (ML) and rule-based techniques.
• The goal is to process large-scale radiology report datasets.

Main Methods:

• A ML algorithm was trained on 1,040 computed tomography pulmonary angiogram reports.
• The pipeline was validated on 49,611 radiology reports from the Mass General Brigham (MGB) healthcare system.
• Performance metrics included accuracy, sensitivity, specificity, PPV, and NPV.

Main Results:

• The initial ML model achieved 91% accuracy on test data.
• Deployment on the MGB dataset showed 85% accuracy, which improved to 94.8% after applying the rule-based algorithm.
• The final hybrid model demonstrated high sensitivity (96.4%) and specificity (93.2%) on the MGB dataset.

Conclusions:

• The hybrid NLP approach requires less training data than pure ML models.
• It demonstrated high performance across diverse healthcare settings.
• This pipeline can efficiently identify PE cases for research and clinical surveillance.