Accurate risk stratification is essential to guide clinical decision-making and enable comparison between surgeons and institutions. Current risk prediction models based on logistic regression show good discrimination but are poorly calibrated for emergency and other high-risk cardiac surgery. Better risk prediction may be achieved by combining risk models into a two-stage approach, modifying existing risk models with additional variables to create disease-specific tools, e.g. for acute aortic dissection, or novel machine learning models. Current machine learning models offer a slight improvement in performance over established logistic regression models, but further improvements and refinements are likely to achieve even better reliability.