Artificial intelligence classification of pediatric middle ear effusion using consumer-grade otoscopes.

INTRODUCTION

Early and accurate diagnosis of middle ear effusion in primary care settings enables timely intervention, improving long-term patient outcomes while preventing overtreatment with antibiotics. Diagnosis requires assessment of tympanic membrane mobility, but this is hampered by access to and inexperience with pneumatic otoscopy or tympanometry. This research aims to train an artificial intelligence algorithm, using tympanic membrane images captured with consumer-grade otoscopes, to accurately predict the presence or absence of pediatric middle ear effusions.

METHODS

Using an over-the-counter digital otoscope, tympanic membrane images were obtained from patients aged 6 months to 6 years old undergoing tympanostomy tube placement at 4 different institutions. The images were labeled according to the middle ear findings identified by pediatric otolaryngologists during myringotomy. A deep learning neural network was trained to classify tympanic membrane images as "Effusion" or "No Effusion".

RESULTS

From 219 patients, 537 tympanic membrane images were obtained. Serous, mucoid or purulent effusion were represented in 263 images and 273 images demonstrated no effusion. The trained model achieved a classification accuracy of 92.1 %, a sensitivity of 90.3 %, and specificity of 93.8 %.

DISCUSSION: This model has a balanced sensitivity and specificity comparable to or surpassing pneumatic otoscopy (94 % sensitivity, 80 % specificity) and tympanometry (90-94 % sensitivity, 50-75 % specificity). With current barriers to using the diagnostic tools recommended in clinical guidelines, it is evident that an inexpensive, objective diagnostic aid is needed for the primary care setting. This model's performance demonstrates the first step toward an accessible tool for primary care providers.