AI tool takes leads from ECGs to help diagnose heart conditions

In a study published on August 1, 2024, in npj Digital Medicine, the scientists from Scripps Research in California report that their AI algorithm can recreate full 12-lead ECGs with data from only three ECG leads.

Moreover, clinicians can identify heart attacks with nearly the same accuracy when they review the AI-generated ECGs as compared to original 12-lead ECGs, which are complex arrangements of electrodes and wires placed around the chest and limbs to detect the heart’s electrical activity.

In a statement, co-senior author and cardiologist Evan Muse, MD, PhD said: “This opens up the door to patients being able to get really high-quality, time-sensitive clinical data without travelling to somewhere that has a 12-lead ECG. It likely means not only increased access to ECG technology, but decreased costs and improved patient safety.”

To build the new AI tool, the team used data from over 600,000 12-lead ECGs that had been collected from patients. About half of those ECGs had normal health rhythms, while the rest had a variety of heart conditions. 

MORE FROM MEDICAL & HEALTHCARE

Giorgio Quer, PhD, co-senior author of the paper, then began testing which combinations of two or three electrodes could be used for AI to fully recreate the 12-lead data.

“We knew that leads are somehow related. Deep learning algorithms allowed us to process a very large dataset and understand these relationships among the leads, enabling the reconstruction of the full 12-lead,” said Quer. “We started out hoping to get a full reconstruction from just limb leads, because those are the easiest for non-specialists to set up. But we found that we got much better data when we added a chest lead as well.”

According to Scripps, the researchers then took a set of 238 ECGs, with half showing signs of a heart attack. They showed cardiologists either the original 12-lead ECG or an ECG reconstructed by AI using data from the three selected leads. The cardiologists could not identify which was which, and they also correctly identified heart attack indicators 81.4 per cent of the time in the AI-generated ECGs, which is very close to the 84.6 per cent accuracy of original 12-lead ECGs.

“It was important to us that we not only show that this algorithm works at a technical level, but that the data generated by the algorithm can be accurately interpreted by cardiologists,” said Quer.

The researchers said that before the algorithm can be used for clinical decision making, prospective studies will be needed with different patient populations and in different clinical settings. However, if the tool continues to perform well, it could lead to ECGs being carried out in new settings with less specialised equipment and clinicians