Otoscopy is a routine clinical examination for evaluating the external auditory canal, tympanic membrane, and middle ear. In primary care and pediatric settings, clinicians commonly perform otoscopy despite variable training and experience, and diagnostic accuracy can therefore vary widely. Visual inspection alone may be insufficient to distinguish conditions that share overlapping or subtle visual features, which can lead to inappropriate treatment or delayed care.
Automated interpretation using deep learning has shown promise for otoscopic images and videos, but many existing models rely principally on visual information and are developed on single-center datasets. In clinical practice, otoscopy is frequently paired with tympanometry, a non-invasive test that measures the mechanical properties of the tympanic membrane and middle ear system. Because tympanometry provides objective physiological data, it can help differentiate visually ambiguous conditions such as effusion versus a normal middle ear.
This study proposes a hybrid, clinically interpretable approach—OtoTymp-AI—that integrates anatomical video information from otoscopy with physiological measures from tympanometry using a confidence-guided fusion strategy rather than joint multimodal representation learning.
OtoTymp-AI is a two-step decision-fusion framework. First, a convolutional neural network (CNN) is trained to classify otoscopy videos across predefined diagnostic categories. Second, the system applies a clinically interpretable, rule-guided decision strategy that uses the CNN’s prediction confidence to determine whether to accept the visual diagnosis or to refine it with tympanometric information.
Specifically, when the CNN yields a high-confidence prediction, that prediction is preserved. When the CNN’s confidence is low, the framework applies tympanometry-based rules reflecting established clinical interpretations of 226 Hz tympanograms to adjust or replace the video-only prediction. The intention of this design is to exploit the complementary information provided by anatomical (video) and physiological (tympanometry) data while keeping the decision process transparent and clinically interpretable.
This was a retrospective multi-center otoscopy study. Three independent institutions contributed data across the development and evaluation pipeline:
The external paired-cohort analysis used 104 videos with paired tympanometry spanning six diagnostic categories. The study emphasized external evaluation because paired otoscopy–tympanometry records are relatively uncommon, and because external cohorts better reflect clinical heterogeneity.
A CNN was trained to classify otoscopy videos into the study’s diagnostic categories. Rather than training a single model to learn a joint video-plus-tympanometry representation, the authors implemented a confidence-triggered fusion: retain high-confidence video predictions and apply rule-based tympanometry logic for low-confidence or uncertain predictions.
The tympanometry rules were designed to be clinically interpretable and to reflect physiologic patterns measured by conventional 226 Hz tympanometry. The threshold for CNN prediction confidence acted as a tunable operating point determining how often tympanometry would overrule or refine the visual prediction. The study explored multiple thresholds and reported outcomes at the best-observed threshold.
In the external paired cohort of 104 paired otoscopy videos and tympanometry, the video-only CNN achieved an overall accuracy of 63.46% (95% CI, 53.88–72.08).
Applying the OtoTymp-AI confidence-guided fusion improved overall accuracy. At the best-observed confidence threshold of 0.90, the hybrid video-plus-tympanometry framework achieved 81.73% accuracy (95% CI, 73.22–87.98). This represented an observed absolute improvement of 18.27 percentage points over the CNN-only model, with a 95% paired bootstrap confidence interval for the improvement of 9.62–26.92.
The study reported that improvements were seen in selected diagnostic categories represented in the cohort, including effusion and retraction. However, per-class performance estimates had wide confidence intervals where sample sizes were small, limiting precise conclusions about some diagnostic groups.
The findings indicate that integrating anatomical video data from otoscopy with physiological data from tympanometry using a confidence-guided, rule-based fusion can increase diagnostic accuracy in an external paired cohort. The approach preserves clinical interpretability by limiting tympanometric intervention to visually uncertain cases and by using explicit, clinically grounded rules rather than opaque joint representations.
Limitations acknowledged by the study include the retrospective design, the relatively small external paired cohort (104 paired records), and wide confidence intervals for some per-class results due to small sample sizes. The authors emphasize that the optimal operating threshold and per-class performance characteristics require confirmation in larger prospective paired multimodal cohorts.
Data and reproducibility resources: a public reproducibility package with code and anonymized derived data is available on GitHub. Raw otoscopy videos, tympanometry records, and associated clinical metadata are restricted due to human-participant protections and may be requested through the responsible institutional channels subject to IRB and data-use agreements.
Overall, OtoTymp-AI demonstrates a proof-of-concept that combining visual and physiological inputs with a confidence-guided fusion strategy can improve AI-assisted middle ear diagnosis in an external paired cohort, while maintaining clinical interpretability and modularity of components.