Artificial Intelligence in Electrocardiology for Arrhythmia Diagnosis
論文ID: CJ-22-0229
この記事には本公開記事があります。
詳細
論文ID: CJ-22-0229
Artificial intelligence (AI) and machine learning (ML) are emerging research areas of interest in imaging diagnosis. The deep learning (DL) model is a subset of ML, and has already been utilized in modern technologies such as automatic facial recognition and voice assistants.1 DL models are able to identify patterns from various biomedical datasets for predicting diseases and conditions. In the field of cardiology, electrocardiography (ECG), echocardiography, pressure waveform, intravascular imaging, and of course, computed tomography and magnetic resonance imaging can all be targets of DL-mediated diagnosis and prediction.1 It has been demonstrated that advanced AI methods such as DL neural networks are applicable to ECG. With the use of large datasets of ECGs linked to clinical variables, AI is now able to perform human-like interpretation of ECGs and even more.2 A number of AI studies using ECG data have been reported. For example, DL models using ECG datasets are able to predict future occurrence of arrhythmia such as atrial fibrillation, or even left ventricular dysfunction (Figure).3 Interestingly, AI is able to estimate an individual’s age, sex and race based on ECG alone.4 It is suggested that AI may interpret patterns and morphologies that cannot be discerned by the human eye.
Schematic of deep learning-mediated ECG diagnosis model.
Article p ????
In this issue of the Journal, Nakasone et al5 report a DL model predicting the origins of outflow tract ventricular arrhythmias, whether right-sided or left-sided. They found that the DL model had more accurate prediction value than conventional ECG algorithms. This is a novel attempt to utilize a DL model with ECG datasets to classify the anatomic origin of cardiac arrhythmias. They also succeeded in visualizing the diagnostic contribution of each ECG lead using the gradient-weighted class activation mapping (Grad-CAM) method, and reveal the possible importance of limb leads in predicting the origin of outflow tract ventricular arrhythmias. Although a DL model is essentially a black-box method, a visualizing approach such as Grad-CAM method may be useful to identify ECG data that a DL model focused on relating to the output of interest. The authors conclude that the visualized DL model is convincing and may play a role in understanding pathogenesis.
This study presents a novel strategy using AI not only to diagnose but also to localize the origin of the disease. Fundamentally, the performance of AI depends on an abundance of data. Due to the small number of evaluated subjects in the present study, there is the possibility of over-fitting by the AI of the incorporated ECG data, which should be confirmed in future studies with large-scale data from a variety of patients.
Future PerspectiveAI with DL models may offer novel ECG interpretations and better prediction of heart diseases and therapeutic strategies. Including cardiovascular imaging, enormous amounts of biomedical data can be processed with AI, which may lead to a futuristic precision medicine era. In addition, data-driven information obtained by an explainable AI such as the Grad-CAM method may depict novel features or pathologies of diseases that should be further focused on. Finally, as the amount of biomedical data continues to increase, it is possible that AI will become essential to clinical practice in cardiology.
None declared.
