Toward an Unbiased Deep Learning Classifier of Pediatric Middle Ear Disease

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pediatric Otolaryngology

Background:

• Otitis media is a common pediatric diagnosis, leading to frequent healthcare visits and antibiotic prescriptions.
• Distinguishing between acute otitis media (AOM) and otitis media with effusion (OME) is critical for appropriate antibiotic use but often challenging.
• Current diagnostic methods rely on expert interpretation of tympanic membrane (TM) visualization.

Purpose of the Study:

• To develop and evaluate an artificial intelligence (AI) algorithm for classifying middle ear conditions.
• To utilize images captured by inexpensive, consumer-grade otoscopes for AI training.
• To accurately predict the presence and nature of middle ear effusion in pediatric patients.

Main Methods:

• A prospective, multicenter cohort study involving 219 children (6 months to 10 years) at four US pediatric hospitals.
• Collection of 737 tympanic membrane images using over-the-counter digital otoscopes during myringotomy procedures.
• Training a deep learning algorithm on labeled images, with intraoperative findings serving as the ground truth.

Main Results:

• The AI classification model achieved a high weighted accuracy of 92.5% across diverse image categories.
• Individual category accuracies ranged from 88.4% to 98.8%, demonstrating robust performance.
• The study included a diverse pediatric population, with 42.14% identified as Black, Hispanic, Asian, or other ethnicities.

Conclusions:

• The deep learning model shows significant accuracy in classifying middle ear status in young children.
• Utilizing affordable digital otoscopes and diverse datasets can facilitate real-world clinical applications.
• This AI approach holds promise for enhancing telemedicine capabilities and optimizing antibiotic stewardship in pediatric otitis media management.