STS Advances Use of Artificial Intelligence, Machine Learning in CT Surgery

David M. Shahian, MD

David M. Shahian, MD

The Society is a recognized leader in quality measurement and public reporting of surgical outcomes. Now, STS is taking a leading role in the use of artificial intelligence (AI) and machine learning (ML) to further improve patient care and outcomes in cardiothoracic surgery. 

“STS is ideally positioned for this leadership role because of our long history in traditional statistical modeling approaches, the ongoing collaborations of many STS surgeon investigators with academic AI/ML departments, and our most unique asset—the premier clinical data outcomes registry in health care: the STS National Database,” said David M. Shahian, MD, from Harvard Medical School and Massachusetts General Hospital in Boston. “Modern AI/ML approaches are ‘data hungry,’ and within the Database, we have not only the most (8 million records), but also the highest-quality, audited, and validated clinical data available.” 

In the coming year, the STS National Database will be further enhanced by supplemental data that include sociodemographic indicators, long-term survival, and reoperation information. Collectively, Dr. Shahian said, it will form an unparalleled source of data from millions of cardiothoracic surgical patients that can be used for AI and ML studies.

The role and future direction of AI/ML is the focus of Tuesday’s session, Machine Learning in Prediction of Cardiothoracic Surgery Outcomes, moderated by Dr. Shahian and Joseph A. Dearani, MD, from Mayo Clinic in Rochester, Minnesota. 

Machine Learning in Prediction of Cardiothoracic Surgery Outcomes

7:00 a.m. – 9:00 a.m.
Room 224

Dr. Shahian explained that many aspects of supervised AI and ML (such as binary classification of a patient’s outcome) are similar to those of conventional modeling, including the selection of a population cohort and identification of potential risk factors. Just as in traditional statistical modeling techniques, it’s important to develop independent training and validation sets and to validate model performance both in the original population from which the models were developed, as well as in new populations; this avoids model overfitting and assures their generalizability.  

“Arguably, the biggest difference between traditional and AI/ML modeling techniques is that the former pre-specify the form of the anticipated relationships between input (e.g., risk factors) and output (e.g., outcomes) variables, which often are assumed to be rather simple (e.g., additive),” Dr. Shahian said. “Conversely, the algorithms in AI/ML learn iteratively from the data to which they are exposed, and they are mathematically much more complex. They may involve complex nonlinear relationships and patterns that are not apparent with traditional approaches.” AI/ML also may identify additional risk factors or combinations of factors that were previously unrecognized. Finally, unsupervised AI/ML algorithms may detect associations or groupings, sometimes called clusters, that otherwise would not have been evident.

With the availability of cloud-based data storage and ever-increasing computer processing speeds, the potential applications of AI/ML continue to expand, Dr. Shahian said. However, he noted that claims about the superiority of this technology for certain tasks—such as predicting patient outcomes based on preoperative data—have been premature and sometimes wildly exaggerated. Accordingly, STS has formed a working group of surgeons and AI/ML experts from numerous academic centers to coordinate these investigative efforts in a strategic, thoughtful manner.

Dr. Shahian said he began exploring AI and ML more than 2 decades ago with colleagues from the Massachusetts Institute of Technology in an attempt to improve the prediction of coronary artery bypass grafting outcomes using a multilayer perceptron neural network. Incorporating data from the STS National Database, it was among the earliest studies utilizing AI and ML in health care. He expects Tuesday’s session to demonstrate how far AI and ML have advanced since those early studies, but also how much opportunity there is for continuing investigation.

“We hope that this session stimulates all STS members to delve deeper into AI and ML techniques and think about possible applications in our specialty,” he said.