Can Machine Learning and AI Replace Otoscopy for Diagnosis of Otitis Media?

In this issue of Pediatrics, Crowson et al¹  provide a glimpse to the future for otitis media (OM) diagnosis, perhaps. Data acquisition from 126 normal and 212 abnormal ear images met study quality criteria when children were brought to the operating room for an examination under anesthesia for tympanostomy tube placement for either recurrent acute otitis media (AOM) or otitis media with effusion (OME). To meet the quality criteria, 75% of the surface area of the tympanic membrane (TM) had to be visible and there had to be sufficient resolution to assess major landmarks. The otolaryngologist classified each ear at the time of myringotomy as effusion present or absent. If an effusion was present, the effusion was categorized as either serous, serous/mucoid, mucopurulent, or purulent. However, the deep learning model was used to target only a binary output of no effusion as normal or any effusion or any degree of TM pathology as abnormal. The model achieved a mean image classification accuracy of 83.8% (95% confidence interval: 82.7%–84.8%).

In the hands of an experienced and trained clinician intent on making a correct diagnosis, otoscopy, especially pneumatic otoscopy, always involves cerumen removal to visualize at least 75% of the TM and is highly accurate.²  However, otoscopy and especially pneumatic otoscopy is challenging to do in an uncooperative infant or toddler for whom the incidence of OM is highest. The ear canals are small and often cerumen is partially blocking the view. Most pediatricians are confident they are accurately diagnosing OM and distinguishing AOM, OME, or a retracted TM without middle ear effusion. Indeed, the American Academy of Pediatrics guideline in 2004³  that included “diagnostic uncertainty” as a consideration in therapy decisions was controversial. But when experienced pediatricians were shown videos of otoscopy that included still frame visualization and pneumatic otoscopy with no cerumen blocking the view, diagnostic accuracy varied from 70% for AOM to ∼45% for OME and retracted TM without middle ear effusion.⁴  Pediatric residents performed similarly.⁵ 

To overcome the diagnostic challenges of otoscopy, several new technologies are being developed. The method described by Crowson et al¹  is one such technology, using video imaging of the TM, followed by machine learning using artificial intelligence (AI), to assist in diagnostic classification of OM. Machine learning is defined as the use and development of computer systems that are able to automatically learn and adapt without following explicit instructions by using algorithms and statistical models to analyze and draw inferences from patterns and data. Machine learning is an application of AI. A group led by Hoberman is using an endoscope attached to a smart phone to take images of the TM and then applying an algorithm to distinguish AOM from OME using machine learning.⁶  TM optical coherence tomography (OCT) is the most advanced new technology in clinical testing. OCT is an optical imaging technique enabling a cross-sectional view of the TM and the identification of thickness changes and biofilm attachment to the TM.⁷  OCT has been shown to facilitate noninvasive differentiation of the types of middle ear effusion.⁸  OCT imaging is also amenable to machine learning application.⁹  Importantly, for all these technologies, machine learning is dependent on the data set provided to the process and the label given to the images used for the AI algorithm to use. This is a data-hungry process that begins with labeling some of the images by an expert as diagnostic for the condition of interest. If labeling is 80% accurate against a ground truth, then the best machine learning can do is 80%.

Better diagnosis can lead to less antibiotic use and less frequent referral for tubes.¹⁰  These are worthy outcomes. However, are they important enough to primary care physicians in practice? Pediatricians in practice are cognizant of keeping on time during a busy day. Therefore, confronted with a young child who may be crying or irritable and resistant to staying still for an examination of the TM, combined with a visual block to much of the TM due to cerumen in the auditory canal, the decision often taken is to diagnose an AOM and prescribe an antibiotic. The parent, eager to get home or to work, is satisfied that the time and effort to bring the child in to be seen was worth it.

So I am obliged to ask these questions: Will any new technology to better diagnose OM be effective if removal of wax to visualize 75% of the TM is necessary? How long will the child need to sit still? Will it not be necessary to differentiate purulent from nonpurulent middle ear fluid? Will the technology be cost-effective if it costs more than a prescription for amoxicillin? Will it be necessary to have the technology readout imported to the electronic medical record and become the ground truth for diagnosis such that a diagnosis of AOM with associated antibiotic prescription or referral for tubes due to prolonged OME needs to be justified when the clinician action contradicts the technology recommendation?

AI

artificial intelligence

AOM

acute otitis media

OCT

optical coherence tomography

OME

otitis media with effusion

TM

tympanic membrane