Aims: Artificial intelligence (AI) tools offer a promising solution for the automated detection and diagnosis of middle ear disease, particularly where specialist access is limited. A critical step is validating a model’s generalisability with external datasets from different imaging systems. A convolutional neural network (CNN) deep learning classification algorithm was originally trained on a large dataset of otoscopic images from Aboriginal and Torres Strait Islander children. Our aim was to assess the generalisability of this algorithm by evaluating its diagnostic accuracy on an independent, diverse cohort of test images obtained using rigid endoscopes instead of otoscopes in adults and not used in its original training. Methodology: Otoendoscopic images were prospectively collected by clinicians from 2 tertiary referral centres in Queensland, Australia. Ground truth diagnosis was established by the consensus diagnosis of two independent otolaryngologists. Images were processed sequentially by the CNN, with predictions recorded. Performance was quantitatively evaluated using accuracy, sensitivity, specificity. Results*: A total of XX images were collected for model validation. The CNN successfully classified the independent endoscopic image dataset with an overall accuracy of XX%, sensitivity of YY%, and specificity of ZZ%, indicating good overall discriminative performance. Overall Cohen’s Kappa coefficient (XX) demonstrated substantial agreement with the specialist consensus. This suggests our model is robust enough to interpret visual differences including lighting, field of view, magnification, depth of field, inherent in these two modalities. Conclusion: The convolutional neural network algorithm maintains high diagnostic accuracy when applied to an independent adult cohort using endoscopic images, despite its original training on otoscopic images in a paediatric Indigenous population. *Results pending, full results will be available by time of ASOHNS