br In fact this higher cell resistance is
In fact, this higher cell resistance is not real, since it 1977535-33-3 is only an error between the PS’ parameters and the real ones from the patient. Thus, some repercussions in the system behavior are expected, more specifically, worse following errors for larger val-ues of PS , which happens for PS = 150 and 250%. However, for
PS = 50% the case is distinct, since its computed MAPE is smaller than the one without uncertainty.
This decrease in the MAPE for small values of PS is due to the existence of an offset in the system. Having a small positive value of PS means administrating a slightly higher drug concentration in the patient, which can dissipate the offset and, consequently, improve the reference following.
3.3. Noise and disturbance variation
The addition of uncertainty in the system is also simulated by adding disturbance d and noise n, respectively, to the input and out-
put. The system is going to be tested for three different situations: R1 with n = 0.1 and d = 10−4 , R2 with n = 1 and d = 10−3 , and R3 with n = 103 and d = 1. As expected, by increasing n and d, the quality of the control sys-tem is deteriorated. That is verified with the increase in the MAPE and in the number of false switches for the disturbed situations when compared with R0 (no noise or disturbance) – Table 1.
For the situation R1 , although the MAPE is slightly higher than in R0 , its curves for the selected model are similar – Fig. 13. This event
Fig. 12. Drug concentration and toxicity level depending on PS . The dashed curves in the left plot represent Ci (t) and the solid curves Cg (t).
Fig. 13. Selected patient dynamics for the three different uncertain situations.
indicates that the values of noise and disturbance of the situation R1 are acceptable for keeping a good performance.
On the other hand, for the two other disturbed situations, this conclusion cannot be expressed. For the situation R2 , there is an increase in the MAPE and, additionally, the switch between the models M2c and M3c is not correctly handled. For the situation R3 , besides the achieved higher MAPE, the MMAC system does not react to any changes, except at the first instants. This is exclusively due to the hysteresis condition. With the addition of noise and disturbance of large amplitude to the system, an “offset” with large varying amplitude is added to all signals of the filtered prediction error i , that had small amplitudes. Then, a set of signals with high order amplitudes that are closer to each other is created. Therefore, once the first model is selected, node of Ranvier is preserved until the end of the therapy since any possible switch “stagnates” in the hysteresis condition.
All the tests and simulations shown provide very satisfactory results, not only in terms of treatment efficacy, but also due to the low toxicity levels achieved. In all the situations, without the addi-tion of error in the system, the tumor is eradicated in less than a year and a half, and the toxicity levels are graded between mild and moderate.
The influence of the immunotherapy in the results is another question to be undertaken, since its concentration is always 150 to 1000 times lower than the anti-angiogenesis concentration. Nev-ertheless, there is a synergy between both therapies.
The clustering algorithms have impact on the behavior of the control system, since the location and the number of centroids, influence the system performance. Particularly, with more clusters, and consequently more controllers, the system performance tends
to degrade, since the system does not have the capacity of react-ing with the necessary speed. Therefore, one can conclude that the addition of the model clustering to the MMAC brought a perfor-mance improvement, since the clustering main goal was exactly to reduce the numbers of controllers to operate. Aside from that, the algorithm that showed the best results – k-means with 6 clusters – does not create clusters with a single model, which indicates that the centroids should be positioned in dense areas of models.
In a nutshell, the results allow to create a new perspective in the creation of computer-aided clinical decision systems, being this work an aid for future research where some improvements can be applied. For example, the clusters with a single model can be considered as outliers and not as an important class of patients. Apart from that, different controller banks can be tested, particu-larly, by using identification systems or neuronal networks to make the set of models adaptable in real time to the variations occurred in patient’s organism.