A comprehensive review on ECG diagnosis and report generation using AI: Technologies, challenges, and future research directions

Abstract

An electrocardiogram (ECG) is a simple and non-invasive method for use in cardiology, but its signals can only be analyzed manually by experts, which takes time and can be unreliable. Artificial intelligence (AI) is a pioneering field that has enabled automation in the computer-aided ECG report generation aimed at improving diagnostic efficiency, accuracy, and accessibility. This systematic review provides an overview of the development and advances of the AI-based ECG report generation. This paper presents a comprehensive review of encoder–decoder models, including convolutional neural networks (CNNs) and recurrent neural networks (RNNs), along with recent transformer-based architectures that accurately track the long-range dependency of the ECG signals. The main focus of this review is on large language models (LLMs) used for automatic ECG report generation. This paper reviews methods that use fine-tuned LLMs as well as the LLM-based retrieval augmented generation (RAG) approach for ECG-to-text generation, which enables it to be clinically feasible. RAG leverages an externally verifiable knowledge base, which significantly improves the clinical reliability of the generated ECG reports. The paper also presents a description of public/private data sources and widely used evaluation metrics for ECG diagnosis and report generation using AI. Additionally, the concept of multi-modal data fusion and other current problems, such as heterogeneity in data, the black-box nature of models, and the importance of ethical guidelines, are discussed. In the future, these reliable multi-modal AI-based models can act as an assistant to clinical experts, and make cardiovascular care readily accessible and highly qualitative.