By Marta Rybczynski
An AI model identified three forms of heart dysfunction from routine electrocardiograms (ECGs) and performed nearly as well when researchers reduced the input from 12 leads to one. The finding suggests a single ECG signal, like the kind collected by some wearables, may be enough to flag patients with conditions including hard-to-detect heart failure with preserved ejection fraction (HFpEF) for further evaluation.
The Wake Forest University School of Medicine researcher team developed the model using more than 1 million ECGs from 165,243 patients at Atrium Health Wake Forest Baptist, then applied it to 72,832 ECGs from 42,880 patients at the University of Tennessee Health Science Center. The findings were published in the Journal of the American Heart Association.
“Our AI model can detect various types of heart dysfunction from a simple, single-lead ECG alone,” said corresponding author Oguz Akbilgic, noting that many smartwatches and wearable ECG devices collect the same lead configuration.
HFpEF can be particularly difficult to detect because the heart pumps out a normal proportion of blood even though it does not fill or function normally. Subtle electrical changes linked to early or asymptomatic heart dysfunction can also be difficult to spot on an ECG, while echocardiography is commonly used to assess left ventricular function.
Operationalising AI Co-pilots in Medical Information and Scientific Communications
Explore how pharma medical affairs teams are operationalising AI co-pilots in medical information and scientific communications with insights on governance, compliance, workflows and best. Learn more about this report.
In the external Tennessee cohort, the 12-lead model reached an area under the curve (AUC) of 0.92 for reduced ejection fraction, compared with 0.90 for the single-lead model. For mildly reduced ejection fraction, the AUC figures were 0.76 and 0.75 for the two models, respectively while for HFpEF they models achieved AUCs of 0.73 and 0.74, respectively. The 12-lead model identified 77 per cent of HFpEF cases at 58 per cent specificity.
Models combining the ECG predictions with age, sex, body mass index, heart rate and cardiovascular conditions improved performance within the Wake Forest data, but those gains disappeared in the Tennessee cohort. Clinical-data-only models also lost substantially more accuracy between the two populations, while the ECG-only models were more consistent, including when researchers used just one lead.
Among 8,418 paediatric ECGs, the 12-lead and single-lead models reached AUCs of 0.97 and 0.94, respectively, for reduced ejection fraction. Only 10 children in the cohort had the condition, however, and the researchers said larger studies are needed before drawing conclusions about how well the models work in children.
The single-lead recordings used in the analysis came from standard ECGs rather than smartwatches or other wearable devices, so the study did not test the model on wearable data. HFpEF cases were also identified using ejection fraction and diagnostic codes rather than adjudication, and the researchers said some undiagnosed cases may have been classified as controls, potentially inflating the number of false positives.
The researchers are now testing the model in a family medicine clinic at Atrium Health Wake Forest Baptist to see whether it can help identify patients who need further evaluation and how its use affects care and resource use.
