Artificial intelligence examining ECGs may predict mortality, AF
Click Here to Manage Email Alerts
PHILADELPHIA — Deep neural networks identified potential adverse outcomes and atrial fibrillation from 12-lead ECGs that were originally interpreted as normal, according to new research presented at the American Heart Association Scientific Sessions.
“Applications of machine learning and artificial intelligence techniques to problems in health care are increasingly common, but generally focus on diagnostic problems such as detecting features in an image of classifying a current diagnosis based on present features,” Christopher M. Haggerty, PhD, assistant professor in the department of imaging science and innovation, and Brandon K. Fornwalt, MD, PhD, associate professor and director of the department of imaging science and innovation, both at Geisinger in Danville, Pennsylvania, told Healio. “Few studies have been able to apply machine learning to the task of predicting future events or patient outcomes. This work is among the first to demonstrate proof of concept for predicting a future patient event — 1-year mortality — with good performance based solely on 12-lead electrocardiography data.”
Predicting 1-year mortality
Sushravya M. Raghunath, PhD, math and computational scientist in the department of imaging science and innovation at Geisinger, and colleagues analyzed 1,775,926 12-lead resting ECGs of 397,840 patients from 34 years of archived medical records. The researchers used ECG traces to train a deep neural network, which would be used to predict mortality at 1 year.
The area under the receiver operating characteristic curve was 0.83. Once age and sex were added, the AUC was 0.85, according to the results.
Utilization of the deep neural network was superior to a nonlinear model that was created with 39 ECG measures both with (AUC = 0.81) and without age and sex (AUC = 0.77; P < .001).
Overall performance of the model was high even in the subset of ECGs previously reported as normal by a cardiologist (n = 297,548; AUC = 0.84); the researchers calculated an HR of 6.6 after the ECG (P < .005), according to the results.
After finding good predictive accuracy within a subset of ECGs considered normal, the next step was a masked survey with three cardiologists to determine whether they were able to see features that indicated the risk for mortality. Results of the masked survey showed that cardiologists were unable to detect the patterns captured by the model despite being shown ECGs labeled true positives and true negatives.
“Defining a patient prognosis is a fundamental task in medicine because it forms the basis for subsequent treatment decisions to help improve the prognosis, as needed,” Haggerty and Fornwalt told Healio. “We believe this type of model can be incorporated into clinical workflows and help to inform that assessment of risk and prognosis. For example, in the setting of advanced chronic disease, an accurate assessment of mortality risk — which we envision to be the product of a physician interpreting this mortality risk prediction in the context of other clinical signals, not solely based on an algorithm output alone — can be used to help inform referrals for palliative care, or signal that more aggressive treatment may be warranted.”
Predicting AF
For the second study, the same team of researchers analyzed 1.1 million 12-lead ECGs that did not show AF from 237,060 patients. Seven percent of patients had incident AF. In a subset of 61,142 patients who had ECGs interpreted as normal, 3% had AF.
A multiclass deep convolutional neural network was trained with fivefold cross-validation to predict incident AF at 1 year with 15 ECG traces as an input.
The prediction model had a mean AUC of 0.75. Each unit risk score increase was associated with a 45% increased odds of developing AF within 1 year (OR = 1.45; 95% CI, 1.15-1.66), according to the results.
The AUC was 0.72 for the subset of patients whose ECGs were interpreted as normal, according to the results.
“This suggests that the algorithms are likely identifying subtle but important features that physicians currently may be overlooking during their clinical interpretations,” Haggerty and Fornwalt told Healio.
This performance of potential population screening was linked to a positive predictive value of 0.3 when screening the highest 1% of patients at risk.
“We are currently working on a prospective clinical trial to determine the real-world value of using artificial intelligence in this setting,” Haggerty and Fornwalt said in an interview. “We hope to demonstrate that the algorithms are successfully identifying patients at very high risk of atrial fibrillation and to ultimately show that earlier treatment can reduce the risk of adverse outcomes such as stroke.” – by Darlene Dobkowski
References:
Raghunath SM, et al. Presentation 119. Big data approaches for cardiovascular risk.
Raghunath SM, et al. Presentation MDP106. Is use of artificial intelligence helpful in solving arrhythmia challenges? Both presented at: American Heart Association Scientific Sessions; Nov. 16-18, 2019; Philadelphia.
Disclosures: Raghunath and Haggerty report they are stock shareholders for Merck. Fornwalt reports no relevant financial disclosures. Please see the abstracts for all other authors’ relevant financial disclosures.