Introduction: Creating and maintaining research databases in trauma can be resource intensive. Natural language processing (NLP) may assist by extracting structured data from free text. We aimed to develop an NLP machine-learning algorithm for classifying traumatic injury mechanism using clinical notes.
Materials and methods: History and physical and discharge summary documentation from trauma patients admitted to a level I trauma center (January 1, 2018 - December 31, 2021) were collected. A 60-20-20 train-tune-test split was used. The primary outcome was injury mechanism (penetrating versus nonpenetrating), with ground truth established by trained abstractors. Bag-of-words and term frequency-inverse document frequency were used to create a weighted corpus. Dimensions were reduced using singular value decomposition. Random forest, support vector machine, and logistic regression were employed. Primary evaluation metrics were area under the receiver operating characteristic curve and accuracy.
Results: 6058 clinical notes from 3029 trauma patients were included. Twenty percent (n = 613) had penetrating injuries, and 48% were due to gunshot wounds (n = 294). Models were constructed with 4944 unique terms, forming 30 topics. Area under the receiver operating characteristic curve in the test set was 0.99, 1.00, and 1.00 for the random forest, support vector machine, and logistic regression models. Accuracy was 0.98, 0.98, and 0.98 for these models, respectively. The run time for training and classification was 234 and 0.15 s, respectively.
Conclusions: A simple NLP machine-learning algorithm for classifying injury mechanism in trauma patients had near-perfect discrimination and accuracy. NLP may be an accurate and scalable tool to automate aspects of chart review for trauma database development as a supplement to registry workers.
Keywords: Artificial intelligence; Database; Injury mechanism; Natural language processing.
Copyright © 2026 Elsevier Inc. All rights reserved.