Lost in calculation: the clinical SYNTAX score goes logistic.

Cardiothoracic surgery pioneered the use of risk models in the reporting of operative results of patients referred to myocardial revascularization. Recently, interventional cardiology has rekindled its interest in constructing specific scores for longterm risk prediction of percutaneous coronary intervention (PCI). One of them, dubbed the SYNTAX score, is a successful model with website dissemination and extensive validation, in which the final score is calculated after coronary angiography by summing the scorings assigned to each individual lesion. Inherent limitations of the SYNTAX score, such as the lack of clinical or ischaemia information, have lately prompted its integration into more sophisticated algorithms with increased prognostic accuracy. There are two ways in which a prognostic model may be useful in clinical decision-making. First, it may be collapsed by tertiles or arbitrary cut-offs to classify patients into risk categories. This knowledge assists in defining the threshold for a procedure and contributes to better resource allocation. Secondly, it can be used to estimate the prognosis of individual patients. These methods may be perceived as two ways to look at the same information, but they differ significantly. For instance, consider a category-based score (i.e. the SYNTAX score) which predicts that 19.2% of patients in the high-risk group will be deceased at 5 years from complex PCI. This model implies that among a large population of patients undergoing complex PCI, 19% of those categorized in the high-risk group will be dead after 5 years. Owing to the typical stochastic and time-dependent nature of prognostic models, classification into risk strata is frequently the best that can be achieved. However, while death is a binary outcome (no/yes, 0/1), the predictions are probabilities ranging between these extremes. Therefore, categorizing patients into schemes may work well in distinguishing between high and low risk, but the ability to give a prediction at the individual level is intrinsically limited. In this issue of the Journal, Farooq and colleagues introduce the Logistic Clinical SYNTAX score, an attempt to tailor risk prediction of PCI patients to individual clinical and angiographic characteristics. Patient-level data from seven contemporary trials (n . 6000) were pooled together and logistic regression analyses performed to explore the independent association of 1-year death with a minimum relevant set of prognostic predictors including age, creatinine clearance, left ventricular ejection fraction, and SYNTAX score. Adding to this core model, an extended model encompassing more clinical variables was investigated. The core model was associated with superior discrimination compared with the SYNTAX score in isolation (areas under the curve 0.75 vs. 0.66) and slightly less discrimination than the extended model (0.79) for the outcome of 1-year death, while no relevant differences were noted among the three models for the combined outcome of 1-year major adverse cardiac events. The study was pragmatically complemented by simple additive score charts for the bedside calculation of the core and extended versions of the logistic Clinical SYNTAX score, in which an additional variable (‘SYNTAX-like’) was included to address the need for recalibrating the risk of 1-year death to patients presenting with complex angiographic presentations (i.e. left main and/or three-vessel disease). The authors should be commended for a nice addition to the field of risk stratification, which indeed introduces a novel tool for the personalized characterization of patients undergoing PCI. Their study adds meaningfully to the evolving understanding that prognostic prediction in PCI should encompass both angiographic and clinical variables. However, a number of unsolved questions remain. First, a key aspect of risk prediction is to consider whether a model coming from a derivation data set is generalizable to similar patients from alternative cohorts. A model will have no clinical value unless there is a demonstration that it works acceptably for patients other than those from whose data the model was obtained. This step is called validation. Importantly, validating a