The science of risk models.

An individual's overall cardiovascular risk should guide appropriate therapy and patient management. Several risk assessment scores are available; however, further development of risk algorithms is necessary to account for changes in available treatments and patient lifestyles, to make use of emerging risk factors and more accurate methods for measuring outcomes, and to provide more targeted measurement of risk for different patient subpopulations. When developing a risk model it is important to clearly define the outcome that the risk will predict, the period of follow up, the patient population, and the predictors to be used and how they will be combined. An appropriate statistical model is specified with the aim of finding the weighted combination of the candidate risk factors that best predicts the disease outcome. Stepwise regression is used to systematically search through candidate risk factors to produce a final model with an acceptable number of highly relevant variables. Possible non-linear effects of continuous variables and interactions between variables must be considered. However, the selection of variables requires not just statistical criteria but also clinical, biological and epidemiological judgement. In general, relatively simple, clinically reasonable and easy-to-use models that can be generalized to other settings are preferred to complex mathematical models that fit the sample data perfectly. There is a permanent need for updating cardiovascular risk scores to reflect advances in our clinical knowledge over time and changes in population risk. Development of a risk model requires both statistical expertise and a sound knowledge of the clinical and epidemiological aspects of cardiovascular disease.