OBJECTIVES: to assess whether Naive Bayes Classification could be used to classify injury causes from the Emergency Room (ER) database, because in the Friuli Venezia Giulia Region (Northern Italy) the electronic ER data have never been used to study the epidemiology of injuries, because the proportion of generic "accidental" causes is much higher than that of injuries with a specific cause. DESIGN: application of the Naive Bayes Classification method to the regional ER database. MAIN OUTCOME MEASURES: sensitivity, specificity, positive and negative predictive values, agreement, and the kappa statistic were calculated for the train dataset and the distribution of causes of injury for the test dataset. RESULTS: on 22.248 records with known cause, the classifications assigned by the model agreed moderately (kappa=0.53) with those assigned by ER personnel. The model was then used on 76.660 unclassified cases. Although sensitivity and positive predictive value of the method were generally poor, mainly due to limitations in the ER data, it allowed to estimate for the first time the frequency of specific injury causes in the Region. CONCLUSION: the model was useful to provide the "big picture" of non-fatal injuries in the Region. To improve the collection of injury data at the ER, the options available for injury classification in the ER software are being revised to make categories exhaustive and mutually exclusive.
