Development of an electrocardiographic signal classifier for bundle branch blocks, applying Tiny Machine Learning

cardiovascular diseases are still the pathologies that generate the highest mortality and economic costs globally. In Latin America, low-income populations are the most vulnerable. Singularly, this population has an incidence of endemic diseases that can lead to blocks of the bundle branch of His. The following study seeks to develop a cardiac abnormality detection system using machine learning techniques and microcontrollers with limited resources to benefit populations with limited access to health environments. The Arduino Nano 33 BLE Sense is employed as the hardware platform due to its ARM Cortex M4 processor and support for TensorFlow Lite. An electrocardiogram (ECG) database is processed using oversampling and under-sampling techniques to address class imbalance. Spectral features are extracted using wavelet transforms, and a multilayer neural network is implemented for classification. Two class balancing approaches are compared: oversampling and undersampling. Results indicate notable improvements in the model’s ability to identify instances of minority classes with the oversampling approach, while undersampling may lead to information loss. The system’s performance is evaluated using key metrics such as precision, recall, and F1-Score. Additionally, computational resources required to implement the model on the Arduino Nano 33 BLE Sense are estimated, with an assessment of Flash and RAM consumption. This analysis is essential to ensure the feasibility of implementation on resource-constrained devices. This work contributes to the advancement of early detection of cardiac anomalies in resource-limited settings, with significant implications for healthcare in underserved communities and rural areas.