Implementing ECG-AI in Clinical Care

There is no doubt that artificial intelligence (AI) is transforming cardiology practice.^[1] Recently, many studies have focused on AI applications in electrocardiogram (ECG) analysis. These ECG-AI models not only excel at identifying obvious findings such as rhythm disorders and ST-segment elevation myocardial infarction, but also show a remarkable ability to detect subtle features that may be imperceptible to the human eye. For example, they can uncover subclinical atrial fibrillation patterns and assess biological aging, suggesting a potential role as novel clinical biomarkers.^[2] Numerous AI studies and models related to ECG have been developed, but the key is to identify practical approaches that can be applied effectively in clinical practice and provide meaningful benefits.

In this issue of JACC, Dhingra et al^[3] address this need by developing and validating an ECG-AI model for detecting and predicting structural heart disease (SHD) with the use of a deep learning algorithm. By analyzing 261,228 ECGs from 93,639 unique patients, they developed a model named PRESENT-SHD, capable of identifying SHD such as left ventricular ejection fraction <40%, moderate-to-severe leftside valvular disease, and severe left ventricular hypertrophy. The model was validated across multiple cohorts, demonstrating strong performance. Although previous studies have used ECGs to predict SHD,^[4] this study offers several distinct strengths and provides key considerations for ECG-AI research.

the use of ECG images—a format ubiquitous in clinical practice—rather than raw ECG voltage data. Because 2-dimensional (2D) images are inherently more complex and often require larger data sets and more sophisticated models than 1-dimensional (1D) signals, the authors used a contrastive learning approach, a self-supervised learning technique, and achieved performance similar to traditional 1D signal methods. Building upon their previous image-based ECG AI approach,^[5] the authors extended its validation to an expanded range of ECG formats—including screenshots and smartphone photographs—thereby demonstrating that high accuracy can be maintained even with widely varying image inputs. Such flexibility ensures vendor independence and broader accessibility across different healthcare settings, making it particularly valuable for real-world clinical implementation.^[6]

Although 2D image-based ECG-AI offers certain implementation advantages, it may not be universally superior. Due to the higher dimensionality of 2D image data, larger deep learning models are required to achieve comparable performance, which in turn may increase computational demands and operational costs (including server specifications and energy consumption). Ultimately, choosing between 2D image-based and 1D signal-based approaches depends on the specific clinical setting and available resources.

Another crucial aspect of practical AI implementation is the perception and acceptance among health care providers. This study demonstrated the clinical utility of ECG-AI not only in detecting current cardiac conditions, but also in predicting the future development of SHD. The forward-looking approach has the potential to transform clinical care by enabling the early identification of high-risk individuals, thereby supporting preventive strategies and timely intervention. Furthermore, the authors made a bold and commendable decision to make their model publicly available through a validation website. This transparency allows physicians worldwide to test the AI model firsthand and assess its practical applicability in their specific contexts.^[2] However, several hurdles remain before full implementation can be achieved, particularly the need for prospective validation studies and cost-effectiveness analyses.

In conclusion, the ECG-AI–based PRESENT-SHD prediction model demonstrates advantages over previous ECG-AI studies, which suggests its clinical practicality. This study signals that the time has come to consider implementation science in ECG-AI research. The implementation science of ECG-AI faces a complex balance between clinical utility, robust performance, technical infrastructure, practical feasibility, and clinician adoption (Table 1). While 2D image–based approaches offer certain advantages in terms of accessibility and vendor independence, they also present their own set of challenges related to computational requirements and resource utilization. The success of any ECG-AI implementation will ultimately depend on finding the right balance between these factors while ensuring robust validation and cost-effectiveness.