Prediction is very difficult, especially about the future.

Prognostic risk models are an important clinical tool that can be used for research purposes, monitoring performance and to evaluate the risk of event occurrence on individual patient level. Validation is needed to determine the generalizability of risk predictions tools in various populations and real-life scenarios. Luehr et al. [1, 2] performed an external validation of the German Registry of Acute Aortic Dissection Type A (GERAADA) score that was designed to predict 30-day mortality in patients suffering from acute type A aortic dissection.

The two primary measures used to assess the performance of a risk prediction tool are calibration and discrimination. Calibration (the ability of the prognostic risk model to predict the absolute risk level) was excellent as the actual 30-day mortality observed in the study (15.1%) was nearly identical to the mortality predicted by the GERAADA score (15.7%). Based on this result, the model could be used for monitoring centre performance. However, the observed model discrimination power (the ability of a prognostic risk model to accurately identify patients at high risk of experiencing the event of interest) was rather low with receiver operating characteristics analysis demonstrating an area under the curve value of 0.673 (95% confidence interval 0.595–0.751). A higher area under the curve value indicates increasing model discrimination power and a value above 0.7 is generally considered acceptable. As the authors propose the model to be used as a bed side tool to explain the situation to the patient and the patient’s relatives, this poor discrimination power should raise concerns. Moreover, additional analyses in various subgroups from the GERAADA score demonstrated poorest model calibration for ‘Previous cardiac surgery’, ‘Preoperative ventilation’ and ‘Inotropes at referral’ subgroups. As the latter two variables are not completely objective but are likely influenced by clinical decision-making, this might not come as a surprise.

Luehr et al. are to be congratulated for their efforts to validate the recently designed GERAADA score risk model. We believe that their results have revealed some important limitations of the model, providing important data to further improve its performance in the future.

Conflict of interest: none declared.