Better Prediction of Outcomes in Pediatric and Adult Cardiac Surgery

Objectives: Can a fundamentally novel analytic approach, driven by machine learning methods, for the prediction of outcomes in pediatric and adult cardiac surgery address the limitations of the current risk prediction models, namely: 1) incapability of handling rare procedures, 2) poor absolute risk estimation, and 3) poor predictive accuracy.

Description

According to The Society of Thoracic Surgeons (STS), there have been over 6.6 million cardiac procedures in adults since 1989 and more than 500,000 records procedures in children. Statistical prediction models for outcomes after a cardiac surgery or procedure have been developed by the STS for eight different types of procedures. These predictions are invaluable in the planning of surgical procedures, in the choice of interventional strategy, in counselling patients, in scheduling procedures and allocating resources.

However, these STS prediction models suffer from three substantial limitations. First and most importantly, these risk models were built around ‘common’ procedures, even though cardiac operations are quite heterogeneous. In fact, STS models leave out ~30% of all procedures, including pediatric procedures and the most complex, high-risk procedures (e.g. patients in cardiogenic shock). This ‘blind spot’ is unacceptable as these high-risk populations are those who probably would benefit the most from an outcome prediction tool, which governs whether they should go through an operation or not. Second, since the STS models are based on logistic regression, the absolute risk in patients at extremely high-risk is often underestimated. Third, the predictive performance of the STS models is quite limited (around 0.70-0.75 AUC).

Through this study, we develop a better tool for the prediction of outcomes in pediatric and adult cardiac surgery that addresses the limitations of the current risk prediction models. Specifically, we leverage machine learning methods and two novel analytical paradigms: stacked models and transfer learning to create an algorithm that will accurately predict outcomes in all patients, including patients that are in the ‘blind spot’ of current prediction models: the highest risk patients and pediatric patients.