Digital control arm and digital twins

Digital twins are virtual subjects that are identical in every way to the real patients in the clinical study of interest. This identity concerns the baseline characteristics and must extend to the environment of these patients, first and foremost; the concomitant treatments they receive. Digital twins are just a more specific form of the synthetic digital or virtual control group. So we will address both topics in this blog.

In a single arm trial, after observing the evolution of the group of patients treated with the treatment of interest, the problem of the analysis that will allow us to draw from the trial the information we are looking for, namely the efficacy of the treatment with its two dimensions: a qualitative dimension, is the treatment effective? And a two stage quantitative dimension, what is the intensity of its efficacy on this group of patients and on similar patients?

The pivot of the solution is obviously the information we have got from the trial, namely the evolution of this treated group, which in its simplest expression is the difference between the final state and the initial state of the patients.

The problem is none other than that of the causal relationship between the administration of the treatment of interest and the evolution of the group. Because of the play of the confounding factors, an after-before comparison cannot accurately address this essential question.

Another way to get around the problem is to split it into two questions: would this evolution have been different if the patients in the group did not receive this treatment, but a placebo or a competing treatment? And how different would it be?

To answer these questions, we recourse to a synthetic group of patients who serve as a control. A group designed in such a way that the comparison of trends is not inherently biased and is based only on the difference in treatment. The former condition is met if the two groups are similar at baseline, either patient by patient (twins of sorts) or only on average. The latter is met if the control is managed exactly as the real arm was or will be.

Whatever the solution adopted, the constitution of this virtual control group must address an unavoidable and sometimes little-known difficulty: what we know about the patients in the trial arm is only a partial description (see blog n° 11). Some of the descriptors of these patients are not measured, often because they are not measurable, and therefore do not appear in the baseline characteristics on which to rely for building the control. And in this hidden part, certain characteristics may be essential for the effect or lack of effect of the treatment, thus for the comparability of the two groups and the management of bias in the comparison. But we do not know. Potential confounding factors are therefore potentially hidden, some of which are probably unknown. These missing baseline characteristics are of concern for those chasing biases. 

This difficulty also concerns adjustment methods, whether statistical or more qualitative such as the propensity score. In addition to this limitation of their ability to reduce the bias of the comparison, there is another, more theoretical, limit, which is the poor adaptability of the rather rigid structure of these adjustment algorithms to the constitutive flexibility of these complex systems that are living systems.

The various methods available to us to constitute the virtual control group must therefore be evaluated against this criterion: are the hidden patient characteristics adequately managed?

The first method, the most traditional, is manual. It consists of selecting from the literature or from clinical or epidemiological databases a group of patients who have not received the treatment of interest and who look similar to patients in the real arm. Individual or aggregated data could represent the control group. Historical controls fall into this category. The aggregate characteristics of this group are compared to that of the real arm through a statistical test, with the fundamental difficulty that this test is designed to highlight a difference and not an identity (the equivalence test approach is rarely, if ever, used in this situation). With individual data, algorithms exist that can pick up in a data-base patients close to the real patients. With the difficulty of scoring and assessing the closeness.  Another difficulty in both cases is that the quality of the data is very often unverifiable. But the major and unavoidable obstacle is these unmeasured descriptors which, possibly, hide confounding factors.

The other two methods are more recent.

Standard Artificial intelligence . This term covers various techniques that can be used: 1) either to constitute a virtual population which will form or from which will be drawn a control group and whose characteristics are as close as possible to those of the trial arm, with, again the problem of assessing the closeness; 2) or seek to create true digital twins. 

Whatever the intrinsic finesse of the AI technique, it comes up against the obstacle of hidden descriptors. And let’s not forget the problem of the quality of the data that was used to feed the algorithm of the chosen AI technique. 

Causal Artificial Inteligence. A third option is to computationally model the disease and the concomitant treatments received by the patients in the trial in order to be able to simulate the evolution of the control patients, which will serve as a reference for establishing the causal relationship and answering questions about the efficacy of the treatment of interest. 

Provided that the models are sufficiently inclusive and of appropriate granularity, this approach has the advantage over the other two of the ability to take into account those baseline characteristics that are not measurable in humans, although they are sources of confounding factors (hidden descriptors). Of course, they must be included in the available knowledge on the pathophysiology of the disease and the mechanism of action of the treatment of interest. Such a validated model can simulate the course of the condition with the concomitant treatments and without the experimental one. The construction of the model enables us to identify the hidden descriptors. Providing one can assume or derive their distributions, one can create “full” digital twins or a virtual population from which virtual patients close to the real ones (see above) can be drawn. In addition, with little more work, one can simulate other competitive treatments, helping to better position the new treatment.