Identification of patients with carotid stenosis using natural language processing

Area of Science:

• Medical informatics
• Natural Language Processing
• Cardiovascular imaging analysis

Background:

• Medical reports possess structured data suitable for automated patient identification.
• Retrospective patient identification is crucial for clinical research and follow-up.

Purpose of the Study:

• To develop a natural language processing (NLP) model for identifying patients with carotid stenosis (CS) history and presence.
• To utilize ultrasound reports for large-scale, automated patient retrieval.

Main Methods:

• Developed an NLP model using linear classifiers, Convolutional Neural Networks (CNN), and Recurrent Neural Networks (RNN) with attention.
• Processed 1527 ultrasound reports (1220 training, 307 testing) from January 2016 to December 2017.
• Employed bag-of-n-grams and TF-IDF features with logistic regression as a baseline.

Main Results:

• RNN-attention models demonstrated significantly higher specificity and F1 scores for predicting CS history compared to logistic regression.
• All models achieved over 93% accuracy in predicting the presence of CS, with RNN-attention reaching 95.4%.
• RNN-attention models showed statistically significant improvements in specificity over logistic regression.

Conclusions:

• NLP models, including linear classifiers and neural networks, effectively predict carotid stenosis history and presence with over 90% accuracy.
• Advanced neural networks (CNN, RNN-attention) outperform traditional linear classifiers, especially with enhanced features.
• NLP offers an efficient and accurate method for large-scale retrospective patient identification, supporting long-term patient follow-up and research.