SUPPLEMENTAL DIGITAL CONTENT. APPENDIX 2 (Sensitivity

advertisement
SUPPLEMENTAL DIGITAL CONTENT. APPENDIX 2
(Sensitivity Analysis)
Measurement errors for laboratory assays used in this study were CV of 5, 2.4, and 10% for total
cortisol, albumin, and corticosteroid binding globulin (CBG), respectively. We determined the
typical CSRmax and free cortisol half-life values for each of the three clinical groups (control,
sepsis, and SS) as the median values reported in the paper. Using these median parameters for
CSRmax and free cortisol half-life we computed a typical cortisol concentration time course using
forward differential equations solver. We then permuted the typical concentration data (for
cortisol at each time point as well as baseline CBG and albumin) by – 1, 0, and + 1 standard
deviations according to reported assay CVs. This provides a good representation of the
distribution of measurement errors for relevant laboratory values.
This resulted in 35 = 243 simulation databases, for which independent solutions of CSRmax and
FCHL were then generated by least-squares approach using the Levenberg-Marquardt
differential equations solution algorithm, as described in Appendix 1 (Supplemental Digital
Content 1, http://links.lww.com/CCM/B108). The percent errors were computed by subtracting
the target parameter estimates from the corresponding parameter estimates obtained using
simulated (permuted) concentration databases. The resulting percent error distribution in
computed parameters is summarized by its standard deviation. These standard deviations were
25.2, 31.1, and 36.7% for CSRmax for control, sepsis, and septic shock groups, respectively. For
free cortisol half-life, standard deviations were 24.3, 26.6, and 30.1% for control, sepsis, and
septic shock groups, respectively.
The significant between-group differences in CSRmax and free cortisol half-life observed
in this study are not directly affected by the potential propagation of laboratory measurement
error, since these errors are already included in the data used. However, we also note that this
degree of error may limit clinical applications of computed parameters. The use of more accurate
laboratory measurements (i.e. smaller CV) or more frequent cortisol sampling would decrease
the magnitude of error in computed parameters. This conclusion is supported by additional
sensitivity analyses using the permutation and simulation approach as above. For example, we
found that for septic shock group, standard deviations of percent error were reduced from 36.7 to
27.3% and to 20.7% for cortisol sampling intervals of 30, 15, and 10 min, respectively.
Download