Computer Classification of Injury Narratives Using a Fuzzy Bayes Approach: Improving the Model

摘要

This paper summarizes improvements to an earlier developed Fuzzy Bayes approach for assigning coding categories to injury narratives randomly extracted from a large U.S. Insurer. Improvements to the model included: adding sequenced words as predictors and removing common subsets prior to calculation of word strengths. Removing subsets and adding word sequences improved prediction strengths for sequences found frequently in the training dataset, and resulted in more intuitive predictions and increased prediction strengths. Improved accuracy was found for several categories that had proved difficult to code in the past. This study also examined the effectiveness of a two-tiered approach, in which narratives were first categorized at the broad level (such as [falls]), before classification at a more refined level (such as [falls from heights].) The overall sensitivity following a two-tiered approach was 79% for predicting classifications at the broad category level and 66% for the more refined prediction categories.