Bayesian methods: a useful tool for classifying injury narratives into cause groups.

Area of Science:

• Occupational Health
• Data Science
• Biostatistics

Background:

• Large administrative databases contain valuable injury narratives.
• Classifying these narratives into event cause groups is crucial for injury prevention.
• Existing classification methods may be time-consuming or less accurate for large datasets.

Purpose of the Study:

• To compare the performance of two Bayesian classification methods: Fuzzy Bayes and Naïve Bayes.
• To evaluate their effectiveness in classifying injury narratives from worker's compensation claims into Bureau of Labor Statistics (BLS) event codes.
• To determine which method offers superior accuracy for large-scale administrative data.

Main Methods:

• A dataset of 14,000 injury narratives was extracted from worker's compensation claims.
• Two expert coders assigned one-digit and two-digit BLS Occupational Injury and Illness Classification event codes.
• Two Bayesian classifiers (Fuzzy and Naïve) were developed using an 11,000-case training set and evaluated on a 3,000-case prediction set.

Main Results:

• Both Fuzzy and Naïve Bayesian methods demonstrated good performance in classifying injury narratives.
• One-digit BLS codes were predicted more accurately than two-digit codes by both models.
• Naïve Bayes showed slightly higher accuracy (e.g., sensitivity 80% vs. 78% for one-digit codes) compared to Fuzzy Bayes.

Conclusions:

• Bayesian methods, particularly Naïve Bayes, show significant promise for classifying injury narratives in large administrative databases.
• These methods offer an efficient and accurate approach to categorizing injury events for improved occupational safety analysis.
• The findings support the utility of automated classification for understanding injury causes in worker's compensation data.