Artificial intelligence-based models demonstrate superior predictive capabilities for postoperative mortality in cardiac surgery compared to traditional tools, though they require external validation.
How do traditional and machine learning-based risk assessment tools perform in predicting postoperative mortality in adult cardiac surgery?
While traditional risk scores for cardiac surgery mortality have limitations in calibration, emerging AI-based models show superior predictive performance but require further validation before clinical implementation.
Risk assessment for immediate mortality is a vital component of the preoperative assessment in elective cardiac surgeries of the adult population. It is generally used to inform consent and plan postoperative care, but can also help identify patients who need preoperative optimization. Risk assessment for open cardiac interventions remains difficult, as an absolute risk assessment tool is still lacking. In this narrative review, we examine recent data on the predictive performance of commonly used risk assessment tools in cardiac surgery and explore missed opportunities to improve predictive performance, including overlooked independent predictors and alternative calculation strategies, such as machine learning. The literature shows that the most popular risk assessment tools are the Parsonnet score, EuroSCORE II, STS-PROM, and ACEF. These have reasonable discriminative capabilities across most populations but occasionally suffer from poor calibration and over- or underprediction. Preoperative inflammation, functional status, physical performance, nutrition, and frailty are potentially relevant clinical factors that could improve mortality prediction modeling using traditional approaches. By far, the largest advancement comes from artificial intelligence-based models that demonstrate superior predictive capabilities utilizing the same predictors. These models are still in development, have not received external validation, are not yet trusted by physicians, and may not be accessible to all institutions due to computing limitations, and thus are not ready for global rollout. Further research in identifying novel predictors of mortality is required, and efforts are needed to validate machine learning models in external cohorts.
Iacobescu et al. (Mon,) conducted a review in Cardiac surgery. Risk assessment tools was evaluated on Postoperative mortality. Artificial intelligence-based models demonstrate superior predictive capabilities for postoperative mortality in cardiac surgery compared to traditional tools, though they require external validation.