Machine Learning May Widen Use of Tympanometry and Improve Hearing Screening for Children

A new study published in the June 15 edition of Ear and Hearing shows that a computer-based system that uses smartphone technology can help untrained individuals in rural areas more accurately screen for hearing problems in children. By using a large dataset of ear measurements, a team of researchers has developed a hybrid deep-learning model that can classify tympanograms—which are graphs of eardrum movement that help identify problems like common middle-ear infections—so lay people can tell if further action is needed. The authors say it is the “first work to demonstrate that automated clinical decision support can make it possible for laypersons to perform screening tympanometry in rural underserved populations with sensitivity similar to a trained audiologist.”

Hearing loss in children can affect their communication, language, and social skills, and the earlier children get hearing help, the more likely they are to reach their full potential. According to the National Institutes of Health, about 5 in 6 children have at least one ear infection by age 3, and ear infections are the most common reason parents bring their child to a doctor. There are several treatment pathways for ear infections, ranging from watchful waiting to antibiotics and ear tubes that keep fluid from building up behind the eardrum. Additionally, children identified with hearing problems can be provided with amplification, special seating, and other accommodations in school.

The most common way doctors check for middle-ear problems is to look at the eardrum using a handheld otoscope. But another more objective way to assess middle-ear function is through tympanometry, which uses sound and variations in air pressure to detect things like fluid behind the eardrum, problems with Eustachian tube function, or a hole in the eardrum. Unfortunately, tympanometry is not typically part of hearing screening programs in schools due to the lack of screening guidelines, cost, complexity, and the lack of decision-making support for layperson use, say the study authors who have also published data about the hearing screening and referral process in rural Alaska.

The study authors recently developed a smartphone-based tympanometer that may represent an important step toward layperson-administered hearing screening. The researchers tested their machine-learning model on data collected from this smartphone-based device, and it worked well—suggesting the system can be used with inexpensive devices in low-resource areas.

Using their model instead of traditional methods, they estimated that an additional 77 cases of preventable hearing loss could have been identified while only requiring 98 additional cases to be assessed further. They note the benefits of preventing hearing loss outweigh the costs of the additional assessments.

Compared to otoscopy and traditional tympanometry—which require extensive training and can be challenging to perform accurately—the researchers found that their automated system was about as effective. This is important for screening programs because it means that untrained individuals can use the system to identify potential hearing problems.

“Nearly 75% of hearing loss in rural areas is preventable,” study coauthor Samantha Kleindienst, AuD, PhD told HearingTracker. “The addition of a lay-friendly cost-effective tool paired with automated machine-learning capabilities will result in a paradigm shift in how we can start to address hearing loss in the areas that need it most.”

While there have been studies on using machine learning to analyze images from otoscopy, there are fewer studies on automatic classification for tympanograms. The researchers also addressed the challenges of using their system in low-resource settings, where noise and artifacts may negatively affect results. A hybrid model was developed that combines a transparent decision tree with deep-learning techniques to make the system more interpretable and clinically relevant. They say this approach allows for uncertainty estimation, robustness, and flexibility in training the system.

Some limitations to the study exist, including the dataset used was collected with only one type of commercial hardware and software in a specific rural population, so future research should include more diverse devices and populations to ensure the system's effectiveness across different scenarios. However, the study suggests that untrained individuals can use an automated system to accurately screen for hearing problems in children using tympanometry. The hybrid deep-learning model developed by the researchers demonstrates promise in integrating accurate and affordable hearing screening into various communities, especially rural areas and/or those with limited resources.

Researchers included Felix Jin, Ouwen Huang, and Mark Palmeri of Duke University, Sarah Morton, Alyssa Platt, Joseph Egger, and Susan Emmett of Duke Global Health Institute in Durham, NC, and Samantha Kleindienst of Norton Sound Health Corporation in Nome, AK and University of Arkansas for Medical Sciences in Little Rock, Ark. The research was funded by the Patient-Centered Outcomes Research Institute and by the Duke Global Health Institute AI Pilot Research Grant.