Figure Legends

Figure 1. Schematic structure of the semi-mechanistic model of Henrich et al. and our comprehensive mechanistic model of thrombocytopoiesis compared in this work. We present major compartments as boxes and cell or TPO fluxes as black arrows. Actions are depicted as thick lines with arrows (stimulation) respectively orthogonal lines (suppression). Different colors correspond to different kinds of actions.
Figure 2. Distribution of degrees of thrombocytopenia observed at different treatment cycles (primary data set of 135 patients from NHL-B trial). Percentages of patients with thrombocytopenia degree 0 is continuously decreasing from cycle to cylce. Percentages of patients with degrees 3 or 4 increase continuously. In particular, degree 4 thrombocytopenias only occur from cycle 3 and increase up to 10.1% at cycle 6.
Figure 3. Comparisons of model predictions and data for the mechanistic and semi-mechanistic model for four selected patients from the NHL-B trial with either nearly periodic behavior (patient #69) or with strong cumulative platelet toxicity during the course of the therapy (patients #814, #1678, #1261). Model performances are comparable for these cases. Compared to the semi-mechanistic model, the mechanistic model performs slightly inferior at the first cycle but better at later cycles 5 and 6).
Figure 4. Comparisons of model predictions and data for the mechanistic and the semi-mechanistic model for eight patients from NHL-B study with irregular platelet dynamics during the course of the therapy with (patients #37, #398, #489) or without clear cumulative toxicity (patients#14, #15 and #677,  #1017, #1428). Both models fail to fit the degree of thrombocytopenia correctly. The mechanistic model performs better for patients #14, #15, #398 and #1017.
Figure 5. Comparison of the predictive powers of the mechanistic and the semi-mechanistic model. Models are either calibrated on a prescribed number of previous cycles (x-axis) or used parameter settings obtained from regression models of baseline covariates from the validation data set (no calibration cycles, yellow). We present the average of the absolute differences between predicted and observed degrees of thrombocytopenia for future cycles (DD characteristics, y-axis). Color-coding corresponds to the predicted cycle. The mechanistic model outperforms the semi-mechanistic model in all situations except for the parameters derived from the regression model showing  similar predictive power. The mechanistic model calibrated on the first cycle even outperforms the semi-mechanistic model calibrated on a higher number of cycles.
Figure 6. Comparison of the predictive powers of the mechanistic and the semi-mechanistic model. Models are calibrated on a prescribed number of cycles (x-axis) or used parameter settings obtained from regression models of baseline covariates from the validation data set (no calibration cycles, yellow). We present the average of the absolute differences between predicted and observed degrees of thrombocytopenia for future cycles, but compared to Figure 5, only differences of more than one degree are counted (LDD characteristics, y-axis). Color-coding corresponds to the predicted cycle. The mechanistic model outperforms the semi-mechanistic model in all situations except for the parameters derived from the regression model showing similar predictive power. The mechanistic model calibrated on the first cycle even outperforms the semi-mechanistic model calibrated on a higher number of cycles for most of the cases. The advantage of the mechanistic model is more pronounced than in Figure 5.
Figure 7. Comparison of sensitivities regarding grade III/IV toxicity of the mechanistic and the semi-mechanistic model. Models are calibrated on a prescribed number of cycles (x-axis). The sensitivity of the mechanistic model is higher.
Figure 8. Comparison of Youden’s J statistics regarding grade III/IV toxicity of the mechanistic and the semi-mechanistic model. Models are calibrated on a prescribed number of cycles (x-axis). Youden’s J statistics are higher throughout indicating higher predictive power.
 

References