Development of a supervised machine learning prediction model to detect otitis media with effusion using smartphone tympanic membrane images

Otitis media with effusion (OME) occurs when fluid builds up in the middle ear without an infection. Studies have shown high misdiagnosis rate. The objective of the study is to use machine learning (ML) to differentiate between OME and normal tympanic membrane (TM) images. This is a prospective case-control study. Patients were divided into case/OME and control/Normal groups based on the consensus assessment of two otolaryngologists and objectively measured with a portable tympanometer. TM images were captured using a smartphone equipped with a video-otoscope. Supervised ML was used to classify the TM images into two categories; Normal and OME. A set of images were used for training the algorithm, while the rest were used to test how accurately the algorithm detected middle ear effusion. A total of 111 TM images were collected, 54 in the control group and 57 in the case group. A subset of these images was used for algorithm training resulting in a sensitivity rate of 96% (95% CI: 81.7–99.3), specificity rate of 81% (95% CI: 63.3–91.8), and an accuracy rate of 89% (95% CI: 77.8–94.8). The Algorithm post-training results indicated a sensitivity of 87% (95% CI: 70.3–94.7), specificity of 74% (95% CI: 55.3–86.8), and an accuracy of 81% (95% CI: 68.7–88.9). The supervised ML model demonstrated promising performance in detecting middle ear effusion from smartphone-captured TM images, highlighting the potential of ML to support OME diagnosis by healthcare professionals.