In order to achieve its aims, the CHIMERA Hub is made up of several technical work packages (sub-projects) and timelines; one data-driven, one model-driven and one machine learning-driven - that each follow on from each other.
Work package 1: Statistical learning from clinical data
We are using intensive care unit data to (a) calibrate an expensive computer model of the human lungs, (b) develop a Bayesian decision support system to address the problem of tissue oxygenation in critically ill patients (c) develop imputation algorithms to address missingness and (d) develop MCMC algorithms and software to deal with multimodality in Bayesian inference and stochastic optimisation.
Work package 2: Iteratively testing and improving biomechanical models
Using the data gathered from stage 1, we will create a model that can classify patients into the relevant decision/ risk categories. For accuracy, we’ll be using different models of varying complexities rejecting the worst-performing models to find the best performing candidate model.
Work package 3: Learning biophysical model structure and parameters with neural networks
Now it comes to whittling the data. The best performing models, using the largest cohort of subjects, will be used with the remaining data being used to validate findings. Deep learning architecture will also be used to bring together all the models so we can classify the patients at greatest risk. During this stage, clinicians will be consulted to see how/which of the risk classifications obtained might best assist clinical decision-making.
After finding the models that work for us, we will combine the model with individual organ components like the cerebral-vascular or renal system which will help inform the learning process.