Virtual Dementia Cohort (VDC)

AETIONOMY proposed to create an own, free, virtual patient cohort, mainly to overcome legal and ethical limitations, but also to get access to a cohort with enriched data and ensuring statistical power.

Our rationale is as follows: Ultimately, all clinical studies are just data. For many of the variables measured for example in ADNI and PPMI we actually know the distribution of values. For other variables, we may quantify our a priori knowledge and assume a Gaussian distribution around its known mean. The data were introduced into forward modelling approaches (integrative cause-and-effect modeling informed by biology) and linked to generative brain network models to create virtually individuals (virtual patients), which produce the entire range of human brain imaging data using related simulations. With the help of these approaches, we generated a VDC cohort:

  • including complete imaging data with structural and functional connectivity’s, an activity by partner AMU which re-uses developments of The Virtual Brain initiative, and
  • simulated clinical features using Bayesian modelling, developped in collaboration between Fraunhofer and UCB Pharma.

Any VDC starts with the analysis and deep understanding of a real-world clinical cohort. An analysis of the variables and their distribution in real-world study data lays the foundation for the data model of the study. Clustering of the data may provide insights into variables that correlate and, as a consequence, indicate dependencies between variables. Since VDCs could contain as many virtual patients as desired, statistical power would not be a limitation at this point. Based on these required ingredients an initial cohort of virtual patients is created.

A complete VDC models the dependency between different variables and makes it possible to represent virtual patients in a way that the virtual patient becomes indistinguishable to a real-world patient. This implies the necessity to carefully validate any VDC and to point out significant deviations of a simulated patient from the statistical distribution of real-world patients. Simple versions of VDCs can be already generated from summary tables as we find them in many publications that describe clinical studies.

Figure: Flowchart of our overall analysis approach for healthy and MCI patients from the ADNI cohort on a high abstraction level

In our case, we demonstrated this first for an Alzheimer’s patient cohort, the ADNI Avatar. Following beneficial scenarios are:

  • Integration of major referential studies into one major VDC meta-cohort.
  • Enrichment of VDCs with additional information
  • Sharing patient-level information without sharing patient data
  • “Patients-like-me” approach

Towards Enabling Virtual Clinical Studies with Longitudinal Bayesian Network Modeling

A multi-scale and heterogeneous data set , namely Alzheimer’s Disease Neuroimaging Initiative (ADNI) data has been used intensively for research purposes.