|
Abstract:
|
There is a high clinical interest in the detection and classification of otitis being one of the most common infections in pediatrics and the leading cause of antibiotic prescriptions. Using daily pediatrician diaries can be crucial. However, analyze them manually is proved to be costly, when feasible. Pedianet contains near seven million pediatric visits from 144 family pediatricians throughout Italy, spanning from 2004 to 2017. The present work purpose an automatic system trained to classify Pedianet's records in six mutually-exclusive categories: non-otitis, otitis, otitis media, acute otitis media (AOM), AOM with tympanic membrane perforation or recurrent AOM. Two independent experts developed a gold standard of 6,531 records (including training, validation, and test). A pediatrician specialized in infectious diseases validated the gold standard. We train six deep-learning (DL) architectures and an ensemble model on their top. The latter reached 95·47% in the balanced F1 score, obtaining performance higher than our experts (max: 94·35%). Our analysis confirmed that DL models could have practical and crucial applications in the differential diagnosis of otitis from free text.
|
