Artificial intelligence and cardiac surgery risk assessment

In this issue, Sinha et al. [1] present a comparison between risk modelling carried out using EuroSCORE II and using machine learning or artificial intelligence techniques. The data used for comparison were from 227 087 operations carried out over >7 years and sourced from the substantial UK national adult cardiac surgery database. The study is well conceived and executed. It appropriately divided the database into 2 sets: 1 for development (70%) and 1 for validation (30%) and assessed the calibration and discrimination of the machine learning model against EuroSCORE II [2]. Most readers will know that risk models are assessed using 2 criteria: calibration (how close is the predicted outcome to the actual outcome observed?) and discrimination (how good is the model as distinguishing between a high-risk and a low-risk patient?). The main findings of this study were that, as regards calibration, EuroSCORE II broadly overestimates risk in the UK cardiac surgery population. The study also found that machine learning techniques were associated with a modest improvement in risk model discrimination, with an area under ROC curve of around 0.834 compared with around 0.817 for the original EuroSCORE II model.