The Impact of ECG Compression on Deep Learning-Based Classification

Abstract

Edge-based ECG devices and real-time continuous ECG monitoring can play an important role in reducing cardiovascular disease (CVD)-related deaths. With recent advancements in artificial intelligence (AI), there is a growing need to deploy arrhythmia classification models on edge devices, especially in resource-constrained environments such as wearable monitors and home-based ECG patches. However, ECG signals are high-frequency time-series data, and they generate large volumes of information even in short recordings. Continuous monitoring can easily result in hundreds of megabytes to gigabytes of data per day, making it difficult to store, transmit, or process such data on low-power or low-bandwidth systems. Deep learning models require substantial computational resources when processing high-dimensional raw data, which many edge devices, such as microcontrollers and ARM Cortex-M CPUs, cannot support efficiently. As a result, inference times increase, and the system may fail to deliver timely alerts or classification results. To address these limitations, adopting ECG data compression techniques is crucial. However, ECG signal compression always carries the risk of losing substantial, clinically significant data. Therefore, after compression, the machine learning-based classification models might fail to classify accurately. It is important to determine the exact impact of ECG signal compression on machine learning-based classification models. In this thesis, we investigated the effect of different ECG compressions on machine learning-based ECG arrhythmia classification models.