Natural language processing of head CT reports to identify intracranial mass effect: CTIME algorithm

Area of Science:

• Medical informatics
• Natural Language Processing (NLP)
• Machine Learning

Background:

• The Mortality Probability Model (MPM) is crucial for adjusting illness severity and informing triage decisions.
• Automated MPM use is limited by the need for manual extraction of "intracranial mass effect" (IME) from electronic health records (EHR).
• This study addresses the challenge of automatically identifying IME from clinical text.

Purpose of the Study:

• To develop and validate a natural language processing (NLP) algorithm for detecting intracranial mass effect (IME) in head CT reports.
• To enable automated severity adjustment and triage decision-making within the Mortality Probability Model (MPM).

Main Methods:

• Utilized a dataset of adult ICU head CT reports from 2013-2016.
• Employed independent human labelers to categorize reports for IME presence (yes/no).
• Trained an XGBoost machine learning model using TF-IDF, Word2Vec, and Universal Sentence Encoder vectorization strategies to identify IME from preprocessed text.

Main Results:

• The TF-IDF vectorization with XGBoost achieved a high Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.9625.
• Classified reports into "positive" (LR=24.59), "inconclusive" (LR=0.99), and "negative" (LR=0.05) categories for IME.
• Achieved high accuracy, with only 2.0% false negatives and 8.6% false positives in the test set.

Conclusions:

• The developed NLP algorithm accurately identifies intracranial mass effect (IME) from head CT reports.
• This automation facilitates the use of the Mortality Probability Model (MPM) with electronic health record (EHR) data.
• The findings support the integration of NLP for enhanced clinical decision support systems.