Identifying Burns in Medical Device Reports

Abstract

As part of ongoing post-market surveillance, the US Food and Drug Administration (FDA) accepts medical device reports (MDRs) describing possible device malfunctions. FDA analysts examine these reports individually, assessing whether a widespread device malfunction may be endangering public health. This work offers a classifier that identifies reports likely to involve a person being burned. The classifier uses a set of n-grams, taken from a large body of medical device report texts, as features; the problem’s dimensionality is reduced using a principal component analysis. Three classifiers were then validated using a smaller body of reports annotated by a set of FDA analysts, and the best classifier was selected from the group. The results indicate that a simple support vector classifier with features based on raw n-gram frequencies is able to identify these serious burn cases with 96% recall.