TinyML-Based System for Cost-Effective ECG Rhythm Classification on Embedded Device

Machine learning is expected to significantly improve medical care by utilizing algorithms that detect cardiac arrhythmia. This study focuses on developing a TinyML-based system that conducts real-time local monitoring and classification of five-class electrocardiography (ECG) rhythms using a microcontroller. The system incorporates pre-processing techniques, including ECG signal extraction and P-, R-, and S-wave feature extraction, combined with a 2-input deep learning model that uses large CNN kernels and DNN architecture. The experimental results for the MIT-BIH dataset with the quantized model achieved an accuracy of approximately 97.34% and an F1-score of 0.96, with a model size of 143 KB. The prototype device was capable of real-time operation and tracked the number of abnormal ECG signals over extended periods. Furthermore, it logs the occurrence of abnormal signals, which can be used for further analyses.