This study presents a TinyML-based machine learning system for multi-class ear condition classification on an edge device, designed to process images captured from a digital otoscope camera. Utilizing a dataset comprising five categories, Normal, Acute Otitis Media (AOM), Cerumen Impaction (CI), Chronic Otitis Media (COM), and Myringosclerosis (MYS), the proposed system performs near real-time classification directly on a resource-constrained microcontroller. The model was developed and optimized using the Edge Impulse platform and deployed as a quantized TinyML library. The system architecture incorporates lightweight convolutional neural networks (CNNs) and the EON™ Compiler to ensure efficient memory usage while maintaining high diagnostic performance. Experimental results demonstrate a validation accuracy of 97.5% and a testing accuracy of 96.31%, with a peak RAM footprint of only 240.3K and an inferencing latency of 1482 ms. These findings highlight the potential of TinyML for portable, low-power medical applications, providing a foundation for privacy-preserving, GDPR-compliant, on-device diagnostics in auditory healthcare without reliance on cloud infrastructure.
Serkan Dişlitaş (Wed,) studied this question.