Generating Tiny Deep Neural Networks for ECG Classification on Micro-Controllers

Artificial Intelligence (AI) enabledembedded devices are becoming increasingly important in the field of healthcarewhere such devices are utilized to assist physicians, clinicians, and surgeonsin their diagnosis, therapy planning, and rehabilitation. In current practice,designing such models requires experts in DNN design, healthcare, and embeddedsystems. Additionally, the task of migrating such models to a differentmicrocontroller (MCU) platform typically requires significant effort to re-sizeand/or re-train the model. This paper shows that Neural Architecture Search(NAS) can be used to generate tiny but accurate multi-objective models forclassifying ECG signals. To the best of our knowledge, our framework is thefirst of its kind, for automatically generating a DNN for screening AtrialFibrillation on an MCU. Moreover, our research shows that the proposed NASfinds more accurate tiny models than human-designed ones, and is effective inenabling customized solutions for a resource-limited target platform.