Researchers at Mount Sinai have recently developed an innovative artificial intelligence (AI) model called HeartBEiT for analyzing electrocardiograms (ECGs). This model interprets ECGs as language, which enhances the accuracy and effectiveness of ECG-related diagnoses, especially for cardiac conditions with limited available data for training.
HeartBEiT is specifically designed for ECG analysis and has shown comparable or better performance than other methods while using significantly less data. This makes ECG-based diagnosis more feasible, particularly for rare conditions that affect a smaller number of patients and have limited data for analysis.
Traditionally, physicians rely on visual inspection to identify disease patterns in ECGs. However, for conditions without established diagnostic criteria or with subtle and chaotic patterns, human interpretation can be challenging. Artificial intelligence, including convolutional neural networks (CNNs), has played a transformative role in this field. However, the researchers at Mount Sinai took a different approach by utilizing an image-generating model to create discrete representations of specific parts of the ECG. This allows for the analysis of ECGs as language.
HeartBEiT understands the relationships between these representations and leverages this understanding to perform diagnostic tasks more effectively. The researchers evaluated the model on three specific tasks: determining the presence of a heart attack, identifying the genetic disorder hypertrophic cardiomyopathy, and assessing heart function. HeartBEiT consistently outperformed other tested methods in these areas.
To train HeartBEiT, the researchers utilized a dataset of 8.5 million ECGs collected from 2.1 million patients over a span of four decades from four hospitals within the Mount Sinai Health System. They compared its performance against standard CNN architectures in the three cardiac diagnostic areas. The study revealed that HeartBEiT achieved significantly higher performance, particularly with smaller sample sizes. Additionally, HeartBEiT provided more precise explanations, highlighting the specific regions of the ECG that contribute to a diagnosis, such as a heart attack. In contrast, CNN explanations were often more vague, even when they correctly identified a diagnosis.
Mount Sinai Researchers Use New Deep Learning Approach to Enable Analysis of Electrocardiograms as Language | Mount Sinai - New York
