Machine learning and natural language processing methods to identify ischemic stroke, acuity and location from
Charlene Jennifer Ong1,2,3,4, Agni Orfanoudaki4, Rebecca Zhang4
1Boston University School of Medicine, Boston, Massachusetts, United States of America.
Plos One
|June 20, 2020
Summary
Natural Language Processing (NLP) and Machine Learning (ML) methods accurately extract clinical stroke information from radiographic text. These advanced techniques improve identification, location, and acuity assessment of ischemic stroke in large datasets.
Area of Science:
- Medical Informatics
- Artificial Intelligence in Medicine
- Radiology Informatics
Background:
- Accurate clinical stroke information extraction from unstructured text is crucial for patient care and research.
- Current coding systems like ICD-9/10 may misclassify stroke events, lacking detail on acuity or location.
- Automated data extraction can significantly enhance stroke identification in large datasets, clinical report triaging, and quality improvement initiatives.
Purpose of the Study:
- To develop and evaluate a comprehensive framework for Natural Language Processing (NLP) and Machine Learning (ML) methods.
- To determine the performance of simple and complex stroke-specific NLP/ML techniques in identifying the presence, location, and acuity of ischemic stroke from radiographic reports.
- To compare different NLP approaches, including word embeddings and classification algorithms, for stroke feature extraction.
Main Methods:
- Collected 60,564 Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) radiology reports from 17,864 patients across two academic medical centers.
- Utilized standard text featurization techniques and developed neurovascular-specific word GloVe embeddings.
- Trained and validated binary classification algorithms (including Recurrent Neural Networks and Logistic Regression) on expert-labeled reports for stroke presence, location, and acuity.
Main Results:
- Recurrent Neural Networks with GloVe embeddings achieved the highest performance internally (AUCs of 0.96, 0.98, 0.93 for presence, location, acuity).
- Simpler Bag of Words (BOW) with Logistic Regression showed strong performance for identifying ischemic stroke (AUC 0.95), MCA location (AUC 0.96), and acuity (AUC 0.90).
- GloVe/Recurrent Neural Networks demonstrated better generalization on an external test set (AUCs 0.92, 0.89, 0.93) compared to BOW/Logistic Regression (AUCs 0.89, 0.86, 0.80).
Conclusions:
- NLP/ML methods provide a robust framework for accurately discriminating stroke features from unstructured radiographic text.
- These techniques can effectively identify stroke presence, location, and acuity, offering significant advantages over traditional coding methods.
- The findings suggest NLP/ML tools are valuable for analyzing large data cohorts in both clinical and research settings for stroke investigations.
