A Survey on Deep Learning-Driven ECG-Based Diagnostics: Architectures, Classification Strategies, and Multi-Stage CNN Approaches for Cardiovascular Disease Detection

Cardiovascular disease (CVD) is the leading cause of death worldwide. A rapid, accurate, and non-invasive detection method utilizing Electrocardiography (ECG) is needed; however, it is typically the manual diagnostics that result in human error. The purpose of this paper is to investigate some of the most recent research developments in automated cardiac diagnostics utilizing Deep Learning (DL), particularly Convolutional Neural Networks (CNNs), with regards to ECG data. We propose and evaluate a Two-Stage CNN diagnostic system enabled by a compact IoT acquisition chain of AD8232/Arduino technology to facilitate a lowinference cost transfer learning classification solution using MobileNetV2. The Two-Stage CNN diagnostic algorithm initially performs a binary classification (Normal/Abnormal) yielding a 93.46% accuracy, and thorough multi-class predictions of MI, Prior MI, and Abnormal Heartbeat are performed accordingly to the binary classification being classified as abnormal, thus facilitating a decrease in computational load for edge-based diagnostic purposes. The review contained herein reviewed CNN DL architectures of 1d CNNs, 2D CNNs, and even hybrids for ECG diagnostics, discussed integration with IoT systems for real-time diagnostic capabilities, and considered research considerations including model explainability (XAI), dataset generalizability, and the need for standardized benchmarks as means to validate ECG DL models. In addition to a discussion of issues relating to dataset generalizability and benchmarking considerations, we propose additional research should include multimodal data fusion, as well as inherently interpretable AI models for successful clinical integration.