Near-miss narratives from the fire service: a Bayesian analysis

Area of Science:

• Occupational safety and health research
• Data science and machine learning applications in surveillance
• Firefighter safety and emergency services industry

Background:

• Narrative text analysis combined with coded surveillance data enhances understanding of injury circumstances.
• Near-miss data are crucial for identifying organizational hazards and enabling proactive risk reduction.
• The National Firefighter Near-Miss Reporting System (NFFNMRS) collects valuable narrative data on incidents.

Purpose of the Study:

• To apply Bayesian autocoding techniques to near-miss narrative data for the first time.
• To develop and evaluate autocoding models for categorizing firefighter injury and near-miss narratives.
• To assess the impact of training set size on model performance and compare predictive abilities.

Main Methods:

• Manually coded NFFNMRS narratives served as a training set for two Bayesian autocoding models: Fuzzy and Naïve.
• Model performance was evaluated using sensitivity, specificity, and positive predictive value.
• Cross-validation was performed on a subset of predictions to ensure accuracy.

Main Results:

• The Fuzzy model outperformed the Naïve model, achieving a sensitivity of 0.74 versus 0.678.
• Model sensitivity increased with larger training set sizes, indicating a learning effect.
• Injury records were predicted with higher sensitivity than near-miss records.

Conclusions:

• Bayesian autocoding methods can effectively code both near-misses and injuries from lengthy narratives.
• This approach enables the creation of new quantitative data elements for injury outcome and mechanism.
• Autocoding offers a valuable tool for enhancing occupational safety surveillance in the fire services.