View PDF Version - RePub - Erasmus Universiteit Rotterdam

INTRODUCTION
Dynamic prediction of response using longitudinal profi les 81
Dynamic updating of an individuals prediction of a specific outcome based on new
gathered information is not routinely implemented in prediction models, but can be
of great importance for the individual patient, for the clinician and the further choice
of treatment. We shall introduce the general problem in the setting of therapeutic
treatment of chronic hepatitis B.
In the last years treatment options for chronic hepatitis B have largely been extended.
Still the virus is very difficult to eliminate and only 10-36% of the treated patients
remains in remission after therapy. 1,2 Peg-interferon (PEG-IFN) has proven effective,
but also has its limitations regarding multiple and possible serious side-effects. 3
However, it has been demonstrated that the response to a course of interferon is
durable and leads to both improved survival and reduction of the incidence of hepatocellular
carcinoma. 3,4 Therefore, prediction of response to PEG-IFN is of great
importance. During therapy the patient is monitored at frequently scheduled followup
visits and several markers are measured to anticipate continuation. Different
patterns of these markers have been described, ter Borg et al. 5 , and also flares
(sudden increase) of markers have been identified as possible predictors of response,
Flink et al. 6 Up to now however, these measurements have not been implemented
routinely to update the individual prediction of response. These may even proof
helpful in guiding and supporting the patient through the long treatment and identify
patients who will have little benefit by continuation of treatment. The above
sketched situation is the inspiration of this paper.
Our aim is to develop a dynamic prediction model to update the prognosis of the
individual patient dependent on the new information that becomes available after
each follow-up visit. The statistical challenge is to find a powerful tool to make use
of the multivariate longitudinal profiles to predict the binary outcome variable. We
introduce a new method and elaborate on existing methods. The performance of
the different methods are furthermore extensively compared. Two different views
are considered: (1) directly modeling the prediction of the outcome variable with
the use of logistic regression techniques with repeated updates and (2) indirectly
classifying individuals into an outcome category over time using Bayes’ theorem.
For the direct approach (1) either the observed marker value or the subject specific
pattern of the marker is used as predictor. We suggest a GEE solution directly
entering the observed marker values. 7 In contrast, Maruyama et al. 8 first fitted a
linear mixed effect model to obtain the subject specific patterns of the longitudinal
markers and then entered the estimated random effects in a logistic regression.
Since the estimated random effects have measurement errors an adjustment of the