Prognostic models score

A prognostic model is the mathematical combination of two or more patient or disease characteristics to predict outcome. Confusingly, prognostic models are also termed prognostic indexes, risk scores, probability models, risk stratification schemes or clinical prediction rules (Reilly and Evans 2006). To be useful, they must be shown to predict clinically relevant outcomes reliably. They must, therefore, be derived from a representative cohort in which outcome has been measured accurately. Next, they must be validated, not just in the data from which they were derived (internal validation) but also on data from independent cohorts (external validation) (Wyatt and Altman 1995 Justice et al. 1999 Altman and Royston 2000). Lastly, a model must be simple to use and have clinical credibility, otherwise it is unlikely to be taken up in routine clinical practice (Table 14.1). 

A prognostic model or score must always be independently validated. Simply because a model seems to include appropriately modeled powerful predictors does not mean that it will necessarily validate well, because associations might just occur by chance, and predictors may not be as powerful as they appear. 

External validation of a model means determining whether it performs well in groups of patients other than those on whom it was derived that is, how is it likely to do in real clinical practice. These other groups almost certainly will differ in case mix, referral patterns, treatment protocols, methods of measurement of variables and definition of outcomes. Nevertheless, if a prognostic model includes powerful predictive variables, appropriately modeled, it should vaUdate reasonably well in other groups of patients. For example. Fig. 14.1 shows the vaUdation of the ABCD score on pooled individual patient data from six independent groups of patients with TIA (Johnston et al. 2007) (Ch. 15). 

Christensen E (2004) Prognostic models including the Child-Pugh, MELD and Mayo risk scores - where are we and where should we go. Hepatol 41 334-350. 

On the other hand, a potential problem with simple risk scores is that they may not use the full information from the prognostic variables (Christensen 1987 Royston et al. 2006). If continuous predictors such as age are dichotomized (e.g. old versus young), power is usually reduced (Altman and Royston 2000). Furthermore, if the dichotomy is data derived at the point where it looks best, it may also compromise the generalizability of the score. However, although some loss of prognostic power is almost inevitable, simple scores often perform almost as well as more complex models. One reason for this is that a simple score based on a small number of highly predictive variables is much less likely to be overfitted than a complex score with additional weakly predictive variables and interaction terms. 

---