Electrocardiogram (ECG) anomaly detection plays a critical role in the early diagnosis and timely treatment of patients with heart disease. This is particularly important for medical devices such as implantable cardioverter defibrillators (ICD), where it is essential to have detection algorithms that are not only highly accurate but also consume minimal power. Traditional detection methods often face challenges in achieving optimal performance on embedded devices with limited resources, which necessitates the exploration of new and innovative solutions. In this paper, we present the design of an ECG anomaly detection device that leverages the capabilities of artificial intelligence algorithms. For this purpose, the STM32F303K8T6 development board has been selected as the platform for model design and deployment. Additionally, we have implemented an evaluation system that thoroughly tests and assesses the performance of the deployed algorithms. The study undertakes a comprehensive comparison between deep learning algorithms and traditional machine learning algorithms, culminating in the proposal of a novel classification algorithm that combines hybrid features. Experimental results highlight that this algorithm not only demonstrates superior memory efficiency and reduced latency on resource-constrained ICD devices but also exhibits higher precision and robust generalization capabilities. This research provides a novel and effective solution for ECG anomaly detection, offering significant implications for the design of medical devices and the optimization of detection algorithms.