Artificial intelligence (AI) has proven to be more accurate than doctors in diagnosing pediatric ear infections, according to a groundbreaking study. The AI model, known as OtoDx, demonstrated over 95% accuracy in diagnosing ear infections in a test set of 22 images, surpassing the 65% accuracy rate among a group of 389 clinicians who reviewed the same images. The results were published in Otolaryngology-Head and Neck Surgery.
Pediatric ear infections are common but often misdiagnosed, leading to delays in treatment or unnecessary antibiotic prescriptions. By outperforming human clinicians, the AI model, developed using deep learning techniques, offers the potential to supplement the expertise of doctors and improve the accuracy of diagnoses. This technology could prove invaluable in assisting pediatricians and urgent care clinics with patient evaluations.
Ear infections occur when bacteria accumulate in the middle ear. It is estimated that at least five out of six children in the United States have experienced an ear infection before the age of three. Left untreated, ear infections can result in hearing loss, developmental delays, and even life-threatening complications. However, overtreating children who do not have an ear infection can lead to antibiotic resistance, posing a significant public health concern.
Traditionally, the diagnostic accuracy of ear infections in children based on physical exams has been below 70%. The challenging nature of evaluating a distressed or crying child during an examination, along with the limited experience of many doctors and urgent care providers in ear evaluations, contributes to this lower-than-desired diagnostic rate.
To address this challenge, researchers at Mass Eye and Ear collaborated to develop an AI model capable of accurately diagnosing ear infections. The model was trained using high-resolution photographs of tympanic membranes (eardrums) collected from children undergoing surgery at the hospital. This dataset represented a reliable ground truth compared to AI models that rely on images sourced from search engines. In a previous proof-of-concept study, the model achieved an 84% accuracy rate in distinguishing between normal and abnormal middle ears.
The latest study involved comparing the refined AI model against clinicians. Using more than 639 images of tympanic membranes from children, the model achieved a diagnostic accuracy of over 80.8%. In comparison, the average diagnostic score among 39 clinicians who reviewed 22 new images was 65%. Pediatricians and family medicine/general internists correctly categorized 60.1% and 59.1% of the images, respectively.
The researchers believe that the AI model, although not intended to replace human judgment, could significantly enhance clinicians’ confidence in their treatment decisions. By providing additional diagnostic support, this technology may help reduce misdiagnoses, improve patient outcomes, and minimize unnecessary antibiotic prescriptions.
Early and accurate diagnosis of pediatric ear infections is crucial to ensure the best possible outcomes for children. With further development and refinement, AI-based diagnostic tools like OtoDx could become valuable assets in clinics, aiding in the identification of ear infections and informing clinical decision-making. The researchers hope that this technology will contribute to better healthcare for children and advance the field of pediatric otolaryngology.
