A development has emerged from Cedars-Sinai Medical Center in Los Angeles, California. Between January 1, 2015, and December 31, 2019, a deep-learning algorithm, capable of analyzing electrocardiogram (ECG) waveform signals, was developed to predict postoperative mortality. This innovative model utlized the power of artificial intelligence to uncover hidden risk markers in preoperative ECGs, providing a more accurate prognosis than ever before. It signifies a leap forward in preoperative risk assessments, traditionally limited in their ability to accurately identify the risk for postoperative mortality.
The derivation cohort for this model comprised preoperative patients with available ECGs. These patients were randomly divided into groups for training, internal validation, and final algorithm testing. The model’s performance was astonishing, achieving an area under the receiver operating characteristic curve (AUC) value of 0.83 in the held-out internal test cohort. This surpassed the discrimination power of the established Revised Cardiac Risk Index (RCRI) score, which only managed an AUC of 0.67. The model showed consistent results across two external U.S. health-care systems, marking a large advancement in preoperative care.
The dataset included 45,969 patients with available ECG waveform images, corresponding to 59,975 inpatient procedures and 112,794 ECGs. In the internal test cohort, the deep-learning algorithm’s ability to discriminate mortality risk was evident, with high-risk patients identified by the model exhibiting an unadjusted odds ratio of 8.83 for postoperative mortality. This was in contrast to the RCRI scores of more than 2, which only showed an unadjusted odds ratio of 2.08. The algorithm’s versatility was further demonstrated as it performed equally well across various types of procedures, including cardiac surgery, non-cardiac surgery, and catheterization or endoscopy suite procedures.
This deep-learning algorithm, termed PreOpNet, was trained on waveform signals from single preoperative 12-lead ECGs. It marked a step in the utilization of AI in medical diagnostics, demonstrating an ability to interpret complex medical data in ways that traditional methods could not. PreOpNet’s architecture was designed to be lightweight, efficient, and capable of running on standard computers. It integrated convolutional layers with an inverted residual structure, allowing for the integration of information across ECG leads. The input was a 12-lead ECG obtained within 30 days before an operative procedure, and the outputs were hospitalization-level outcomes following that procedure. This approach enabled a comprehensive evaluation of perioperative ECGs, offering a more nuanced understanding of postoperative mortality risks. The model’s external validation was conducted in two separate health-care systems without any additional tuning or training, ensuring the rigour of the validation process. At the Stanford Healthcare (SHC) cohort and the Columbia University Medical Center (CUMC) cohort, the model’s AUC for postoperative mortality was calculated based on analyses of a single preoperative ECG. This external validation illustrates the model’s robustness and applicability across different healthcare settings.
The development of PreOpNet represents an advancement in perioperative care and risk assessment. By leveraging deep-learning algorithms to analyze preoperative ECGs, medical professionals can now identify postoperative mortality risks with greater accuracy. This model provides clinicians with additional information, aiding them in making informed decisions about medical procedures and future complications. Funded by the National Heart, Lung, and Blood Institute, this study has opened new avenues in medical research and patient care, showcasing the potential of AI in enhancing healthcare outcomes.