NEW YORK, NY (June 6, 2023) – Mount Sinai researchers have developed a groundbreaking artificial intelligence (AI) model for electrocardiogram (ECG) analysis that can interpret the electrocardiogram (ECG) as language . This approach can increase the accuracy and validity of ECG-related diagnoses, especially for cardiology where training data are limited.
In a study published online in the June 6 issue of npj Digital Medicine DOI: 10.1038/s41746-023-00840-9, the researchers found that a new deep learning model known as HeartBEiT provides a foundation on which specialized diagnostic models can be built. was reported to form Created. The research team noted that the model created using HeartBEiT outperformed his established ECG analysis method in comparative tests.
“Our model consistently outperformed convolutional neural networks.” [CNNs], is a machine learning algorithm commonly used for computer vision tasks. Such CNNs are often pre-trained using publicly available images of real-world objects,” says lead author Data-Driven and Digital Medicine (D3M) at Icahn School of Medicine, Mount Sinai. Lecturer Akil Vide, M.D., said: “Because HeartBEiT is ECG specific, it can perform as well or better than these methods, using 1/10th of the data. For rare diseases with limited
Over 100 million ECGs are performed each year in the United States alone, thanks to its low cost, non-invasiveness, and broad applicability in heart disease. Nonetheless, physicians are unable to consistently identify patterns that describe disease with the naked eye, particularly in situations where there are no established diagnostic criteria or where such patterns are subtle or chaotic for human interpretation. For conditions, the usefulness of the ECG is limited in scope. However, artificial intelligence is currently revolutionizing science, and most of the research so far has centered around his CNN.
Based on his keen interest in so-called generative AI systems such as ChatGPT, Mount Sinai is taking the field in bold new directions. The system is built on Transformers, a deep learning model trained on large text datasets to generate human-like text. Respond to user prompts on almost any topic. The researchers used a related image generation model to create a discrete representation of a small portion of his ECG, allowing the ECG to be analyzed as language.
“These representations can be considered individual words, and the entire ECG can be considered one document,” explains Dr. Vaid. “HeartBEiT understands the relationships between these representations and uses this understanding to more effectively perform downstream diagnostic tasks. whether you have hypertrophic cardiomyopathy, whether you have a genetic disorder called hypertrophic cardiomyopathy, and how efficiently your heart is functioning. It performed better than all other baselines tested.”
Researchers pretrained HeartBEiT using 8.5 million ECGs from 2.1 million patients collected over 40 years from four hospitals within the Mount Sinai Medical System. We then tested its performance against standard CNN architectures for three cardiac diagnostic domains. In this study, we found that HeartBEiT performed significantly better even with small sample sizes and had better “explainability”. Senior Author Girish Nadkarni, M.D., MPH, Eileen and Arthur M. Fishberg, Ph.D., Icahn Mount Sinai Professor of Medicine, Director of the Charles Bronfman Institute for Personalized Medicine, Office of Data-Driven and Digital Medicine The system manager of the department) explains in detail. M.D.: “Neural networks are thought of as black boxes, but our model is more specific to the ECG regions involved in diagnosing heart attacks and other conditions, allowing clinicians to gain a deeper understanding of the underlying pathology. By comparison, CNN’s description was vague, even if the diagnosis was pinpointed accurately.”
In fact, the Mount Sinai team has greatly enhanced the ways and opportunities a physician can interact with an ECG through a sophisticated new modeling architecture. “We want to make it clear that artificial intelligence is in no way a replacement for ECG-based expert diagnosis,” explained Dr. Nadkarni. health. “
The title of this paper is “Fundamental Vision Transformers Improve ECG Diagnostic Performance.”
This study was funded by the NIH National Heart, Lung, and Blood Institute (Grant No. R01HL155915) and the NIH National Center for Promotion of Translational Sciences (Grant No. UL1TR004419).
To view a complete list of authors and competing interests, see DOI: 10.1038/s41746-023-00840-9.
