Risk prediction models for diagnostic or prognostic outcomes are useful tools for clinical decision support. Most commonly, a dichotomous outcome (e.g. a benign or malignant tumor) is considered. Especially in diagnostic problems, however, a differential diagnosis often includes more levels than categorization of subjects as diseased versus non-diseased (e.g. a benign, borderline or invasive tumor). Methods for updating existing risk prediction models, i.e. adjusting an existing model in order to improve predictions from future patients in a new and different setting, had already been suggested for dichotomous models but did not yet exist for multinomial models. Closely related, the aspect of calibration of multinomial risk prediction models, i.e. the reliability of the predicted risks, had not been studied extensively. Therefore, in this dissertation we extended calibration statistics, calibration plots as well as updating techniques to prediction models for polytomous outcomes based on multinomial logistic regression.