On validation of cardiovascular risk scores.

In predictive modeling, as reviewed by these authors, historical data are used to predict a known future. Each model produces a risk score on a scale of 0-1 (or 0-100%), but each outcome is either an event or a non-event, so the score cannot ever be “valid”, except in an extreme situation. As in many of the papers reviewed, typically the basic definition of “validation” is stretched to include directional agreement of observed relative frequencies and predicted risks across ordinal categories of predicted risk, but there remains an underlying problem that, when any score is used in clinical practice, we are predicting an unknown future. How can we possibly validate this?