Chapter 23Detecting Middle Ear Disease
—Yixi Xu and Al-Rahim Habib
Executive Summary
Hearing loss affects about 1.5 billion people—or about 20 percent of the global population—and it is anticipated to impact over 2.5 billion people by 2050. The economic burden of hearing loss nears $1 trillion in the United States alone. Globally, about 34 million children suffer from hearing loss; preventable hearing loss in children includes chronic middle ear infection that can be detected by otoscopy (looking into the ear with a magnification device). We sought to evaluate the generalizability of artificial intelligence algorithms that use deep learning methods to identify middle ear disease from otoscopic images. To do so, we used 1,842 normal and abnormal otoscopic images that were collected from three independent sources in Van, Turkey; Santiago, Chile; and Ohio in the United States.
We used deep learning methods to develop models, and we evaluated their internal and external performance using area under the curve (AUC) estimates. Further, we conducted a pooled assessment by combining all three cohorts with five-fold cross-validation. We found that artificial intelligence–informed otoscopy algorithms had high internal performance (mean AUC: 0.95, 95% Confidence Interval (CI): 0.80–1.00) but lower performance when tested on external otoscopic images not used for training (mean AUC: 0.76, 95% CI: 0.61–0.91). Model performance when applied to the combined cohort had better pooled performance ...