Causal AI and probabilistic graphical models

What we do

We combine structural causal models with machine learning to answer interventional questions, such as the expected effect of a treatment in patients with given characteristics, and counterfactual questions, such as how the glucose trajectory of a specific patient would have evolved under a different lifestyle, given what has been observed for that patient. Causal graphs are learned from data and refined with expert knowledge, and can be coupled with mechanistic models so that counterfactual predictions remain consistent with physiology. We also use causal inference to emulate clinical trials from real-world data.

Causal models are a class of probabilistic graphical models, which represent the dependencies among variables with a graph. We develop Bayesian networks and hierarchical Bayesian models that share information across related data sets, such as those collected in different hospitals, cohorts or experimental batches, so that each data set benefits from the others while keeping its own characteristics.

Applications. Personalised prevention of type 2 diabetes, treatment decisions in intensive care, effect of fluid therapy on chest X-rays, multi-centre clinical studies.