AI-interpreted ECGs predict mortality risk after heart surgery
Click Here to Manage Email Alerts
Key takeaways:
- A deep learning algorithm applied to preoperative ECGs improved mortality risk prediction.
- The algorithm ran successfully on a standard clinical workstation.
A novel deep learning algorithm applied to a single preoperative ECG improved risk prediction for death after cardiac surgery and other inpatient procedures for a large cohort, researchers reported.
Compared with a widely used standard perioperative risk assessment tool and alternative ECG assessment tools, the deep learning algorithm, called PreOpNet, was able to more effectively identify high-risk patients who went on to experience postoperative mortality, David Ouyang, MD, a cardiologist and researcher at the Smidt Heart Institute at Cedars-Sinai, and colleagues wrote in The Lancet Digital Health. The researchers validated the findings in two external cohorts with diverse patient populations.
“We have been excited by the opportunity to use artificial intelligence to prognosticate risk in clinically important scenarios given AI’s ability to identify and quantify subclinical disease as well as other data points that clinicians cannot see,” Ouyang told Healio. “Doctors are limited in their ability to predict the risk of postoperative outcomes. If predicted accurately, such risks can make a big impact in the decision to go for surgery as well as modify the operative plan.”
Retrospective ECG data
The researchers developed the deep learning algorithm in a derivation cohort of preoperative patients with available ECGs from Cedars-Sinai Medical Center from 2015 to 2019, randomly splitting patients (8:1:1) into subsets for training, internal validation and final algorithm test analyses. Ouyang and colleagues analyzed data from 45,969 patients with a complete ECG waveform image available for at least one 12-lead ECG performed within 30 days before a procedure date. There were 59,975 inpatient procedures and 112,794 ECGs; these included 36,839 patients in the training dataset, 4,549 in the internal validation dataset, and 4,581 in the internal test dataset. The mean age at the time of preoperative ECG was 65 years; 45.1% were women and 21.7% had preexisting CAD.
“We trained and validated a deep learning algorithm based on waveform signals from a single preoperative 12-lead ECG, termed PreOpNet, on the outcome of postoperative mortality,” the researchers wrote. “The input of the model was a 12-lead ECG obtained within the 30 days before an operative procedure, and the outputs were the hospitalization-level outcomes following that procedure. Patients with multiple procedures were treated independently during model training, with each ECG paired with the most proximal subsequent procedure.”
Researchers assessed model performance using area under the receiver operating characteristic curve (AUROC) values in the hold-out test dataset and in two external hospital cohorts and compared with the established Revised Cardiac Risk Index (RCRI) score. The primary outcome was post-procedural mortality across three health care systems.
Predicting postoperative mortality
In the held-out internal test cohort, the algorithm discriminated mortality with an AUROC value of 0.83 (95% CI, 0.79-0.87), surpassing the discrimination of the RCRI score with an AUROC of 0.67 (95% CI, 0.61-0.72), according to the researchers.
The algorithm also discriminated mortality risks in two independent U.S. health care systems, with AUROCs of 0.79 (95% CI, 0.75-0.83) and 0.75 (95% CI, 0.74-0.76), respectively.
Patients determined to be high risk by the deep learning model had an unadjusted OR for postoperative mortality of 8.83 (95% CI, 5.57-13.2) compared with an unadjusted OR for postoperative mortality of 2.08 (95% CI, 0.77-3.5) for RCRI scores of more than 2.
In analyses stratified by procedure type, the deep learning algorithm showed similar performance for patients undergoing cardiac surgery (AUROC = 0.85; 95% CI, 0.77-0.92), noncardiac surgery (AUROC = 0.83; 95% CI, 0.79-0.88) and catheterization or endoscopy suite procedures (AUROC = 0.76; 95% CI, 0.72-0.81).
“We were pleasantly surprised by how robust the AI algorithm was,” Ouyang told Healio. “Our algorithm, PreOpNet, worked well to predict postoperative mortality across three institutions, including two intuitions where it never saw training data, and across a wide range of types of procedures and surgeries.”
Ouyang said all algorithms should be used cautiously until there is further validation, such as with a prospective clinical trial, and further studies of PreOpNet are ongoing.
“The ECG contains additional hidden information beyond human evaluation, and this data can help risk stratify patients before surgery and procedures,” Ouyang told Healio. “AI can assist physicians in the complex decision-making process to undergo surgery.”
For more information:
David Ouyang, MD, can be reached at david.ouyang@cshs.org.
Collapse
Click Here to Manage Email Alerts