Optimizing Cardiac Surgery Risk Prediction: An Machine Learning Approach with Counterfactual Explanations

Dengkang Qin, Mengxue Liu, Zheng Chen, Qian Lei

Abstract: Postoperative complications after cardiac surgery can be severe and even fatal, making it a high-risk procedure. Predicting surgical risk can guide the effective formulation of treatment plans for high-risk cardiac surgery, thereby reducing the risk of postoperative complications, which has attracted widespread attention from cardiac surgeons. The most commonly used method, EuroSCORE, has the problems of low prediction accuracy and weak targeting for postoperative complications. In this paper, we developed a machine learning (ML) model for predicting adverse outcomes (AO) after cardiac surgery with high accuracy and demonstrated the clinical interpretability of the model with counterfactual explanation (CE) based explainable artificial intelligence (XAI). A total of 2324 patients who had undergone cardiac surgeries with cardiopulmonary bypass support in a single center were included in this study, were divided into two groups as non-AO (n = 2148) and AO (n = 176). Our ML prediction model showed the best prediction performance using perioperative data (AUC = 0.769) when compared with models of EuroSCORE (AUC = 0.663) and EuroSCORE covariates (AUC = 0.710). CE method applied to the ML model showed how operation duration, ASA class, BMI, Lac entering ICU and PLT value increase the risk of adverse outcomes following surgery. In addition, sufficiency and necessity metrics was used to provide CE with a better explanation of feature importance. It has been proven that machine learning models have shown hope in improving the risk assessment of adverse outcomes after cardiac surgery, and counterfactual explanations methods provide more detailed and practical explanations, which are more useful for medical professionals.