“An agent-based model of host response to infection as a proxy system for control discovery using evolutionary computation and game-playing Artificial Intelligence”
Dr. Gary An
Sepsis, which is brought about by the body’s host response to severe infection or injury, is one of the most prevalent causes of mortality in intensive care units (ICUs). Sepsis has a mortality rate of ~30-40% and a cost of more than $20 billion annually in the US. The fundamentals of sepsis care, antibiotics, fluid management and organ support, have not changed in nearly 30 years, and to date there is no single approved drug that targets the pathophysiological processes that drive the host-response that produces sepsis, this despite tens of billions of dollars spent on hundreds of failed clinical trials. We have proposed that the controllability of sepsis can be examined by using a previously validated agent-based model (ABM) of the host response to infection as a proxy model upon which different methods of control discovery have been applied. Specifically, we treat the search for an effective multi-modal treatment regimen as a control-optimization problem that manipulates the internal variables of the ABM with combinations of putative molecular-based interventions at different intervals. Given the combinatorial complexity of the high-dimensional potential control space we have applied both genetic algorithms and deep reinforcement learning (as used in the game-playing DeepMind artificial intelligence systems, e.g. AlphaGo and AlphaZero) to characterize the scale of the control problem. Implemented on high-performance computing environments and following the principle that clinical heterogeneity is a function of model parameter space, both approaches produced fairly generalizable solutions, but with acknowledged limitations in interpretability and potential clinical translation. We suggest that these technologies can be integrated with ABM development in an iterative workflow that can both continually refine the ABM as well as guide basic and translational research in sensor and drug design. This approach for multi-scale model-based control discovery is potentially applicable to any complex disease process.