Naïve Bayes classifiers with feature selection to predict hospitalization and complications due to objects swallowing and ingestion among European children

The inhalation/ingestion of a foreign body in the aerodigestive tract is a serious health problem in paediatric patients. With the increasing availability of large databases with a wide range of information there is a growing need to use methods capable of recognizing classification patterns.Based on a dataset of 356 foreign body injuries located in larynx, pharynx, mouth, esophagus or stomach, which occurred in European children aged up to 14 years in 2000–2002, this study is aimed to predict injury severity applying feature selection procedures for identifying the most influential variables among the consumer product characteristics and the circumstances of the injury that can lead to hospitalization and complications.Selective naïve Bayes classifiers were implemented on a subset of 307 and 303 injuries, for which hospitalization and complications details were given respectively, using filter and wrapper strategies for inducing feature selection.The variables recognized by the selective naïve Bayes were: the injured child age and gender, the foreign body type and its characteristics (volume, consistency, shape), whether the injury occurred in presence or in absence of an adult and the activity the child played before the accident. Filter strategy resulted in the most accurate classification model.Models induced by feature selection approach give an improvement in the comprehensibility of the phenomenon, which is missed when the whole set of variables is considered. The need of fostering the attention of caretakers toward a proper surveillance of children and of the manufactures toward the safety design of the end product emerged.

► Injuries due to foreign body inhalation/ingestion in children 0–14 were considered.
► The most influential variables associated to severe injuries were evaluated.
► Naïve Bayes models with different feature selection strategies were implemented.
► Naïve Bayes with filter strategy resulted in the most accurate classification model.
► The importance of proper surveillance of children and product safety design emerged.