Estimating Nonfatal Gunshot Injury Locations With Natural Language Processing and Machine Learning Models

Area of Science:

• Public Health
• Injury Epidemiology
• Computational Medicine

Background:

• Nonfatal gunshot injuries are a significant public health concern in the US.
• Accurate data on the locations of these injuries is limited, hindering prevention strategies.
• Natural language processing (NLP) offers a potential solution to extract location data from unstructured medical records.

Purpose of the Study:

• To evaluate the effectiveness of NLP and machine learning (ML) models in predicting nonfatal gunshot injury locations.
• To generate updated national estimates of where these injuries occur.
• To inform targeted firearm injury prevention efforts.

Main Methods:

• A cross-sectional study utilized data from the National Electronic Injury Surveillance System Firearm Injury Surveillance Study (1993-2015).
• NLP techniques were employed to generate predictors for injury location from medical text narratives.
• Four ML models (multinomial support vector machines, lasso regression, XgBoost, feed-forward neural networks) were trained and validated to classify injury locations.

Main Results:

• The analysis included over 59,000 unweighted gunshot injury records, weighted to approximate national estimates.
• Existing data suggested homes as the most common injury location, followed by streets/highways.
• NLP-enhanced analysis revealed streets/highways as the most frequent location (46.1%), followed by homes (37.7%).

Conclusions:

• NLP and ML models demonstrate utility in classifying nonfatal gunshot injury locations.
• The study indicates a higher prevalence of injuries occurring on streets/highways compared to previous estimates.
• These findings provide crucial data for developing more effective firearm injury prevention strategies.