Promise and pitfalls of prognostic models for epilepsy surgery

www.thelancet.com/neurology Vol 14 July 2015 683 mentioned by Jehi and colleagues, t h e c o n c o r d a n c e s t a t i s t i c s (c-statistics) for the validation models were close to 0·5, suggesting that the models work only somewhat better than by chance. Were the c-statistics for the development models similarly small? It would be interesting to know whether the c-statistics are low because of diff erences between the cohorts (ie, if the cohorts are dissimilar then the same nomogram might not work as well) rather than because of issues related to the prognostic model design. Certain potentially important predictors were not included. How much of an effect do Jehi and colleagues believe that the short duration of follow up (media n of 3·3 years in the development cohort and 4·0 years in the validation cohort) had on model discrimination? To what extent might the low c-statistics be related to some degree of selection in the validation cohort given that four major centres recruited only 604 individuals in 5 years (a mean of about 30 recruitments per centre per year)? Whether these new prognostic models are any more discriminating than the educated opinion of an experienced clinician remains unclear.