View PDF Version - RePub - Erasmus Universiteit Rotterdam

Chapter 2.3
70
and p=0.056 for the interaction term week*decline HBV DNA log10 when added to model
B. These fi ndings corroborate the increasing OR presented in table 2.
Table 3. Dynamic logistic regression analysis of sustained response. Extension of the baseline model
with on-treatment factors.
Model On-treatment factors Comparison OR 95%CI p-value
A Rx week
+ HBV DNA during Rx
B Rx week
+ decline HBV DNA during Rx
C Rx week
+ ALT during Rx
D Rx week
+ ALT during Rx
1 log10 (copies/ml)
decrease / 4 wk
1 log10 (copies/ml)
1.20 (1.03 ; 1.40) 0.017
decrease / 4 wk
1 xULN (IU/l)
1.28 (1.08 ; 1.51) 0.004
increase / 4 wk
1 log-e xULN (IU/l)
1.06 (0.95 ; 1.19) 0.289
increase / 4 wk 1.18 (0.70 ; 2.00) 0.525
Model A, B, C and D use as an offset the baseline subject specifi c PEG-IFN HBV treatment index
Application of the model in clinical practice
Figure 2 suggests that a cut-off at 2-log10 decline in HBV DNA levels provides optimal
discrimination. The maximal chi-square approach of a grid of cut-off points of HBV DNA
decline (table 4) supported that a 2 log10 decline in HBV DNA within 24 weeks of therapy
resulted in the best prediction. To validate the fi ndings 500 bootstrap samples were
drawn and the cut-off point search was repeated; in 89% of cases the maximal cut point
was 2 log10 decline, in 11% it was 2.5 or 3 log10 decline. The discriminative ability of 2
log10 decline compared well with the dynamic logistic regression of HBV DNA decline
(the c-statistics of 0.867 versus 0.863) and was higher than the baseline model (c-statistic
0.846). When the 2 log10 decline of HBV DNA was used as predictor of response alone,
i.e. without the baseline prediction, a NPV of 94% was obtained. The optimized predicted
probability of SR, dependent on the occurrence or absence of a 2 log10 decline
in HBV DNA within 24 weeks of therapy, is shown in fi gure 4. This optimized probability
of SR can be used to decide whether or not to continue PEG-IFN therapy. For example
a patient with a 30% prediction of SR at baseline with more than a 2 log10 HBV DNA
drop during the fi rst 24 weeks of treatment will have a higher SR prediction rate of 40%,
whereas the same patient without a 2 log10 drop before week 24 will have an updated
SR rate of less than 10%, or a negative predicted value of 90%.
Table 1 shows that the vast majority of patients in this study were infected with genotype
A and D, and only a small proportion of patients harboured HBV genotype B and C. To
test whether the 2 log10 rule applies for all genotypes the maximal chi-square approach
of a grid of cut-off points was repeated including the interaction term with genotype.
In none of the models the interaction term was signifi cant (p>0.71) and the lowest