Detection of arrhythmias from ECG signal using TINYML-based embedded system

In this study, arrhythmia classification of ECG (electrocardiogram) signals using TinyML was performed on the STM32H750B board, utilizing a series of quantization, feature extraction, and deep learning algorithms. Wavelet transform was applied to the ECG signals obtained through segmentation from the MIT-BIH Arrhythmia database, and the extracted features were used to train an LSTM network. MATLAB was used for data preprocessing and model development. The combination of the LSTM model, designed for time series, and the Morlet wavelet function, which contains both time and frequency information of the signal, enabled successful learning of features in both rhythmic and non-rhythmic ECG signals. TinyML technology was employed to optimize the model for execution on a resource-constrained device (STM32H750B-DK). To perform the Morlet wavelet transform calculations on the embedded system, the MATLAB CODER extension was used to recreate the wavelet transform algorithm, which is predefined in the MATLAB library, for the STM32H750B-DK development board. After a series of software optimizations to reduce the memory requirements of the development board, Morlet wavelet transform calculations and machine learning algorithms were successfully executed together on the STM32H750B-DK development board. The primary goal of this study is to develop a model that classifies arrhythmias in ECG data with high accuracy using TinyML on embedded systems. By the end of the study, the accuracy of the model running on the embedded system reached satisfactory levels. This study demonstrates the potential of TinyML not only for physiological signals but also for analyzing complex data on resource-constrained devices.