A team of researchers from Stanford University, working with cardiac monitoring company iRhythm, have created an AI algorithm that, in a small proof-of-concept trial, outperformed board-certified cardiologists at identifying various types of arrhythmias from ECGs. The results have been published to arXiv, an online archive for scientific papers, but have not yet appeared in a peer-reviewed journal.
In the study, researchers obtained a dataset of of 64,121 ECG records from 29,163 patients from iRhythm, all 30-second snippets collected from iRhythm’s Zio Patch. The dataset also contained annotations from cardiologists. Researchers used 90 percent of the data to train a complex neural network.
They used the other 10 percent for the test, running those ECGs through their algorithm as well as showing them to six individual cardiologists. Both results were compared for accuracy against the gold standard: a consensus from a committee of board-certified cardiologists.
The network outperformed the average cardiologist in detecting all but two of the 14 arrhythmias studied, and outperformed them for both sensitivity and specificity.
“One of the big deals about this work, in my opinion, is not just that we do abnormality detection but that we do it with high accuracy across a large number of different types of abnormalities,” Awni Hannun, a graduate student and co-lead author of the paper, told the Stanford News Service. “This is definitely something that you won’t find to this level of accuracy anywhere else.”
While the team is not the first to use AI to interpret ECGs, the project distinguishes itself both in the number of different arrhythmias their algorithm can identify and in the size and nature of the dataset their learning neural network was trained on.
“A recently released dataset captured from the AliveCor ECG monitor contains about 7,000 records,” the researchers write. “These records only have annotations for Atrial Fibrillation; all other arrhythmias are grouped into a single bucket. The dataset we develop contains 29,163 unique patients and 14 classes with hundreds of unique examples for the rarest arrhythmias.”
It’s worth noting that AliveCor’s internal dataset, which it used to develop its FDA-cleared ECG analysis algorithms, may well be much larger than 7,000. The company told MobiHealthNews in 2014 that it had more than a million ECGs it was using to train algorithms.
But AliveCor has mostly focused in on detecting atrial fibrillation, whereas this research shows the algorithm can distinguish between at least 12 different arrhythmias, including some where the differences are subtle. Hannun and his team can see a lot of potential in the technology.
“Given that more than 300 million ECGs are recorded annually, high-accuracy diagnosis from ECG can save expert clinicians and cardiologists considerable time and decrease the number of misdiagnoses,” the paper concludes. “Furthermore, we hope that this technology coupled with low-cost ECG devices enables more widespread use of the ECG as a diagnostic tool in places where access to a cardiologist is difficult.”