The most complex model discussed in this chapter represents marker observations generated by a latent discrete-time stochastic process. This avoids strong assumptions about the shape of individual marker profiles, which may consist of long and irregular series of measurements. We discuss approaches to recursive Monte Carlo updating of predictions, for use in realtime forecasting, and present a method for comparing the predictive ability

a: 2 LU ~ a: 1 • <( ~

FAILURE Figure 18.1 Observations on fictitious patient k, monitored between time O and failure. Dots represent marker observations {which may be missing in any time slot). In our notation, these data are Yk1·= 2.7, Yk2 = 2.3, Yk3 = LO, dk1 = d1<2 = 0, dk3 = I.