Hybrid in silico evaluation of insulin dosing algorithms in diabetes

Abstract Clinical trials are the commonly accepted proof of validity of therapeutic approaches. In the case of type 1 diabetes mellitus (T1DM), the therapeutic approach consists basically in replacing the missing exogenous insulin production by a self-subministration of insulin analogues, following rules fixed by the medical doctor and taking into account several factors, in particular the expected carbohydrate intake. The rules of for insulin dosing are based on experience and on the analysis of glucose values. In view of the large number of possible options, it would be very useful to be able to screen different insulin dosing rules in clinical trials, but unfortunately, the necessary clinical trials would be too expensive and complex to realize, so that only few variants can be really tested, and in no case tailored to the specific patient. Against this background, there has been a substantial interest in using complex physiological models of the human glucose metabolism to estimate the effect of therapeutic approaches in simulation, the so-called in silico evaluations. If the results are consistent for a large cohort of virtual patients, the results will not prove as conclusive as real clinical trials, but are usually accepted to be indicative of the real therapeutic outcome. As an alternative (or extension), recently, several methods have been proposed in the scientific literature which follow a similar idea, but do not rely solely on physiological models, but try to extrapolate the effect of a modified therapy using real measurements as baseline, generating thus a hybrid in silico framework for virtual clinical studies of insulin dosing algorithms. The key idea of all these so-called “Deviation Analysis” methods consists in splitting the measurements of the recorded data into a controllable part (which is described by some kind of model of insulin action), and a “disturbance” component, which is assumed to represent all the unexpected variations due to non recorded effects like movement, psychic state etc. The main expected advantage of such approaches are the inclusion of a realistic level of intrapatient variability in the simulation studies (and therefore the prediction of more realistic glucose trajectories), which is expected to lead to better performance estimates when analyzing glycemic outcomes for a new control strategy on a cohort level. The main limitation is the validity region, i.e. these methods are less reliable if larger changes of the insulin dosing with respect to the one used in the baseline measurements are to be simulated. The current paper gives an overview of the ideas behind Deviation Analysis approaches and presents recent, new ideas for performing such types of hybrid in silico evaluations.