A Dynamic Bayesian Network to Model Dysfunctioning Behavior of Multiple Organ Systems for Sepsis
In the Intensive Care Unit (ICU), Sepsis is associated with high mortality rates and generally affects the old or the very young. In order to fight an infection, sometimes our bodies mount an overwhelmingly systemic immune response and in the process, injures its own tissues and organs. While the host response and infection status at a given time can only be partially observed, dysfunction status of multiple organ systems can be estimated. We currently lack an understanding of how these failing organ systems dynamically interact with each other over the length of ICU stay and if this evolution pattern is disease-specific. Through the use of temporal generative models, we investigate if there is a time-based physiological coupling of organ system states and make predictions for future events in light of new evidence.
We hypothesise that different diseases have characteristic multivariate evolution patterns whose temporal changes can be modelled with a dynamic bayesian network (DBN).
The main tasks of the DBN:
1. Given complete data, search over the graph space to find the most likely network structure to explain the observed data for septic and non-septic patients independently.
2. Learn the parameters of the graph i.e conditional probability distribution of the organ states at the current time with information from previous time step(s).
3. Reason if the learned graph structures can be informative to discriminate the dysfunctioning evolution of Sepsis patients from non-sepsis patients.
4. Finally, evaluate the method on a different patient population.